From 38bd9c9bd16d5feaa692c17b96beb76e3b14653c Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 16:54:25 +0900
Subject: [PATCH 01/11] feat(ask): show cited-post evidence in a Layer Popup

Reading an Ask answer's evidence today means either scanning the inline
fact list or leaving the answer entirely to open the full post popup.
Add AskEvidenceLayerPopup (frontend/src/components/), a focused modal
layer -- opened via a new "View evidence" button per citation -- showing
that post's text evidence facts (checkpoint 3's cited_post_evidence) and
image evidence (checkpoint 3's cited_post_images) without navigating
away from the answer. Proper dialog semantics: role="dialog",
aria-modal, Escape-to-close, backdrop-click-to-close, initial focus on
the panel -- stricter accessibility than the existing PostDetailPopup,
which has none of these.

Extracted chatEvidenceKindLabel into evidenceKindLabels.ts so both
App.tsx and the new component share one label map instead of drifting.

Stacked on #419 (checkpoint 3): the popup's image-evidence section needs
that PR's cited_post_images field to be meaningful.

Part of the Ask Agent temporal/lineage/evidence goal (checkpoint 4 of 4).
---
 frontend/src/App.test.tsx                     |  21 ++++
 frontend/src/App.tsx                          |  37 +++++--
 .../AskEvidenceLayerPopup.stories.tsx         |  79 +++++++++++++
 .../components/AskEvidenceLayerPopup.test.tsx |  85 ++++++++++++++
 .../src/components/AskEvidenceLayerPopup.tsx  | 104 ++++++++++++++++++
 frontend/src/evidenceKindLabels.ts            |  12 ++
 frontend/src/i18n.ts                          |   8 ++
 7 files changed, 336 insertions(+), 10 deletions(-)
 create mode 100644 frontend/src/components/AskEvidenceLayerPopup.stories.tsx
 create mode 100644 frontend/src/components/AskEvidenceLayerPopup.test.tsx
 create mode 100644 frontend/src/components/AskEvidenceLayerPopup.tsx
 create mode 100644 frontend/src/evidenceKindLabels.ts

diff --git a/frontend/src/App.test.tsx b/frontend/src/App.test.tsx
index de8d2821f..8bbf6acce 100644
--- a/frontend/src/App.test.tsx
+++ b/frontend/src/App.test.tsx
@@ -1712,6 +1712,27 @@ describe("App, authenticated", () => {
     expect(screen.queryByText(/Image evidence:/)).not.toBeInTheDocument();
   });
 
+  it("opens a cited post's evidence in a Layer Popup without leaving the answer", async () => {
+    stubBackend({ askImageCitation: true });
+    render(<App />);
+    expect(await screen.findByRole("button", { name: "View post: Public post" })).toBeInTheDocument();
+    await userEvent.click(screen.getByRole("button", { name: "Ask Agent" }));
+    await userEvent.type(screen.getByRole("textbox", { name: "Ask a question" }), "Which project?");
+    await userEvent.click(screen.getByRole("button", { name: "Ask" }));
+
+    await userEvent.click(await screen.findByRole("button", { name: "View evidence" }));
+
+    const dialog = await screen.findByRole("dialog", { name: "Linked post" });
+    expect(dialog).toBeInTheDocument();
+    expect(within(dialog).getByText(/project: Semantic project/)).toBeInTheDocument();
+    expect(within(dialog).getByText("Screenshot of the checkout error")).toBeInTheDocument();
+
+    await userEvent.click(screen.getByRole("button", { name: "Close evidence panel" }));
+    expect(screen.queryByRole("dialog")).not.toBeInTheDocument();
+    // The answer itself is still on screen -- the layer never navigated away.
+    expect(screen.getByRole("button", { name: "View evidence" })).toBeInTheDocument();
+  });
+
   it("labels the Customer Master entity level and Keymen side, never the raw lookup code", async () => {
     // Live UI finding (2026-08-19): read_customer_master() skipped the
     // common_lookup_value join both endpoints elsewhere already use,
diff --git a/frontend/src/App.tsx b/frontend/src/App.tsx
index e5c6d4bee..ccfcaaed4 100644
--- a/frontend/src/App.tsx
+++ b/frontend/src/App.tsx
@@ -85,7 +85,9 @@ import {
 import { CitationChip } from "./components/CitationChip";
 import { CutoffKnownBody } from "./components/CutoffKnownBody";
 import { LineageEntityPicker } from "./components/LineageEntityPicker";
+import { AskEvidenceLayerPopup } from "./components/AskEvidenceLayerPopup";
 import { PopupCloseButton } from "./components/PopupCloseButton";
+import { chatEvidenceKindLabel } from "./evidenceKindLabels";
 import { BuyerNav, type BuyerDestination } from "./components/BuyerNav";
 import { LineageDag } from "./LineageDag";
 import { PostBody } from "./PostBody";
@@ -777,16 +779,6 @@ function projectProvenanceLabel(provenance: string): string {
   return t(PROJECT_PROVENANCE_LABELS[provenance] ?? "Recorded evidence");
 }
 
-const CHAT_EVIDENCE_KIND_LABELS: Record<string, string> = {
-  source_field: "Source field hint",
-  semantic_project: "Semantic project",
-  semantic_role: "Semantic role",
-  semantic_keyman: "Semantic Keyman",
-};
-
-function chatEvidenceKindLabel(kind: string): string {
-  return t(CHAT_EVIDENCE_KIND_LABELS[kind] ?? "Evidence");
-}
 
 const VERIFICATION_BADGE: Record<string, string> = {
   verify_pending: "Not yet checked",
@@ -4477,6 +4469,7 @@ function AskAgentPanel({
   const [answer, setAnswer] = useState<AskAgentResponse | null>(null);
   const [error, setError] = useState<string | null>(null);
   const [asking, setAsking] = useState(false);
+  const [evidenceLayerPostId, setEvidenceLayerPostId] = useState<string | null>(null);
 
   async function handleAsk() {
     const normalized = question.trim();
@@ -4525,6 +4518,13 @@ function AskAgentPanel({
                     <button className="post-list-item" onClick={() => onOpenPost(post.post_id)}>
                       <strong>{post.post_title}</strong>
                     </button>
+                    <button
+                      type="button"
+                      className="citation-chip"
+                      onClick={() => setEvidenceLayerPostId(post.post_id)}
+                    >
+                      {t("View evidence")}
+                    </button>
                     {answer.cited_post_evidence?.find((item) => item.post_id === post.post_id)?.facts.length ? (
                       <ul className="post-evidence-list" aria-label={t("Evidence facts")}>
                         {answer.cited_post_evidence
@@ -4555,6 +4555,23 @@ function AskAgentPanel({
           )}
         </section>
       )}
+      {evidenceLayerPostId && answer ? (
+        <AskEvidenceLayerPopup
+          postId={evidenceLayerPostId}
+          postTitle={
+            answer.cited_posts?.find((post) => post.post_id === evidenceLayerPostId)?.post_title ??
+            evidenceLayerPostId
+          }
+          facts={
+            answer.cited_post_evidence?.find((item) => item.post_id === evidenceLayerPostId)?.facts ?? []
+          }
+          images={
+            answer.cited_post_images?.filter((image) => image.post_id === evidenceLayerPostId) ?? []
+          }
+          onClose={() => setEvidenceLayerPostId(null)}
+          onOpenPost={onOpenPost}
+        />
+      ) : null}
     </section>
   );
 }
diff --git a/frontend/src/components/AskEvidenceLayerPopup.stories.tsx b/frontend/src/components/AskEvidenceLayerPopup.stories.tsx
new file mode 100644
index 000000000..f542f50d1
--- /dev/null
+++ b/frontend/src/components/AskEvidenceLayerPopup.stories.tsx
@@ -0,0 +1,79 @@
+import type { Meta, StoryObj } from "@storybook/react-vite";
+import { AskEvidenceLayerPopup } from "./AskEvidenceLayerPopup";
+
+const meta = {
+  title: "Evidence/AskEvidenceLayerPopup",
+  component: AskEvidenceLayerPopup,
+  args: {
+    postId: "post-demo-public",
+    postTitle: "Checkout error follow-up",
+    facts: [
+      { kind: "semantic_project", text: "project: Checkout revamp | evidence: Body evidence" },
+      { kind: "semantic_keyman", text: "Keyman mention: Ada West | context: account lead" },
+    ],
+    images: [
+      {
+        unit_index: 1,
+        caption: "Screenshot of the checkout error",
+        extracted_text: "Error code 500 on checkout",
+      },
+    ],
+    onClose: () => undefined,
+    onOpenPost: () => undefined,
+  },
+} satisfies Meta<typeof AskEvidenceLayerPopup>;
+
+export default meta;
+
+type Story = StoryObj<typeof meta>;
+
+export const Default: Story = {};
+
+export const TextEvidenceOnly: Story = {
+  args: {
+    images: [],
+  },
+};
+
+export const ImageEvidenceOnly: Story = {
+  args: {
+    facts: [],
+  },
+};
+
+// Edge case: a citation with no persisted evidence facts or images at all --
+// must show an explicit placeholder, never a blank panel.
+export const NoEvidence: Story = {
+  args: {
+    facts: [],
+    images: [],
+  },
+};
+
+// Edge case: an image evidence entry whose OCR text was never extracted.
+export const ImageWithoutExtractedText: Story = {
+  args: {
+    facts: [],
+    images: [
+      {
+        unit_index: 0,
+        caption: "Architecture diagram",
+        extracted_text: null,
+      },
+    ],
+  },
+};
+
+// Edge case: an untitled/uncaptioned image.
+export const UntitledImage: Story = {
+  args: {
+    facts: [],
+    images: [
+      {
+        unit_index: 0,
+        caption: null,
+        extracted_text: null,
+      },
+    ],
+  },
+};
diff --git a/frontend/src/components/AskEvidenceLayerPopup.test.tsx b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
new file mode 100644
index 000000000..314a8ac92
--- /dev/null
+++ b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
@@ -0,0 +1,85 @@
+import { render, screen } from "@testing-library/react";
+import userEvent from "@testing-library/user-event";
+import { describe, expect, it, vi } from "vitest";
+import { AskEvidenceLayerPopup } from "./AskEvidenceLayerPopup";
+
+const baseProps = {
+  postId: "post-demo-public",
+  postTitle: "Checkout error follow-up",
+};
+
+describe("AskEvidenceLayerPopup", () => {
+  it("renders text and image evidence for the cited post", () => {
+    render(
+      <AskEvidenceLayerPopup
+        {...baseProps}
+        facts={[{ kind: "semantic_project", text: "project: Checkout revamp | evidence: Body evidence" }]}
+        images={[
+          {
+            unit_index: 1,
+            caption: "Screenshot of the checkout error",
+            extracted_text: "Error code 500 on checkout",
+          },
+        ]}
+        onClose={vi.fn()}
+        onOpenPost={vi.fn()}
+      />,
+    );
+
+    expect(screen.getByRole("dialog", { name: "Checkout error follow-up" })).toBeInTheDocument();
+    expect(screen.getByText(/project: Checkout revamp/)).toBeInTheDocument();
+    expect(screen.getByText("Screenshot of the checkout error")).toBeInTheDocument();
+    expect(screen.getByText("Error code 500 on checkout")).toBeInTheDocument();
+  });
+
+  it("shows an explicit placeholder when the citation has no persisted evidence", () => {
+    render(
+      <AskEvidenceLayerPopup {...baseProps} facts={[]} images={[]} onClose={vi.fn()} onOpenPost={vi.fn()} />,
+    );
+    expect(
+      screen.getByText("No persisted evidence is available for this citation."),
+    ).toBeInTheDocument();
+  });
+
+  it("closes on backdrop click, close button click, and Escape, but not on panel click", async () => {
+    const onClose = vi.fn();
+    const { container } = render(
+      <AskEvidenceLayerPopup {...baseProps} facts={[]} images={[]} onClose={onClose} onOpenPost={vi.fn()} />,
+    );
+
+    await userEvent.click(screen.getByRole("dialog"));
+    expect(onClose).not.toHaveBeenCalled();
+
+    const backdrop = container.querySelector(".popup-backdrop") as HTMLElement;
+    await userEvent.click(backdrop);
+    expect(onClose).toHaveBeenCalledTimes(1);
+
+    await userEvent.click(screen.getByRole("button", { name: "Close evidence panel" }));
+    expect(onClose).toHaveBeenCalledTimes(2);
+
+    await userEvent.keyboard("{Escape}");
+    expect(onClose).toHaveBeenCalledTimes(3);
+  });
+
+  it("opens the cited post from the layer", async () => {
+    const onOpenPost = vi.fn();
+    render(
+      <AskEvidenceLayerPopup
+        {...baseProps}
+        facts={[]}
+        images={[]}
+        onClose={vi.fn()}
+        onOpenPost={onOpenPost}
+      />,
+    );
+    await userEvent.click(screen.getByRole("button", { name: "Open post: Checkout error follow-up" }));
+    expect(onOpenPost).toHaveBeenCalledWith("post-demo-public");
+  });
+
+  it("moves initial focus onto the dialog panel", () => {
+    render(
+      <AskEvidenceLayerPopup {...baseProps} facts={[]} images={[]} onClose={vi.fn()} onOpenPost={vi.fn()} />,
+    );
+    expect(screen.getByRole("dialog")).toHaveFocus();
+  });
+});
diff --git a/frontend/src/components/AskEvidenceLayerPopup.tsx b/frontend/src/components/AskEvidenceLayerPopup.tsx
new file mode 100644
index 000000000..1b9ca8b76
--- /dev/null
+++ b/frontend/src/components/AskEvidenceLayerPopup.tsx
@@ -0,0 +1,104 @@
+import { useEffect, useId, useRef } from "react";
+import { chatEvidenceKindLabel } from "../evidenceKindLabels";
+import { t, tf } from "../i18n";
+import { PopupCloseButton } from "./PopupCloseButton";
+
+export type AskEvidenceLayerFact = {
+  kind: string;
+  text: string;
+};
+
+export type AskEvidenceLayerImage = {
+  unit_index: number;
+  caption: string | null;
+  extracted_text: string | null;
+};
+
+export type AskEvidenceLayerPopupProps = {
+  postId: string;
+  postTitle: string;
+  facts: AskEvidenceLayerFact[];
+  images: AskEvidenceLayerImage[];
+  onClose: () => void;
+  onOpenPost: (postId: string) => void;
+};
+
+/**
+ * A focused evidence layer for one Ask Agent citation -- opened from the
+ * answer without leaving it, unlike the full post detail popup.
+ *
+ * Next action: open the source post for the complete record, or close to
+ * return to the answer.
+ */
+export function AskEvidenceLayerPopup({
+  postId,
+  postTitle,
+  facts,
+  images,
+  onClose,
+  onOpenPost,
+}: AskEvidenceLayerPopupProps) {
+  const headingId = useId();
+  const panelRef = useRef<HTMLDivElement>(null);
+
+  useEffect(() => {
+    panelRef.current?.focus();
+  }, []);
+
+  useEffect(() => {
+    function handleKeyDown(event: KeyboardEvent) {
+      if (event.key === "Escape") onClose();
+    }
+    document.addEventListener("keydown", handleKeyDown);
+    return () => document.removeEventListener("keydown", handleKeyDown);
+  }, [onClose]);
+
+  return (
+    <div className="popup-backdrop" onClick={onClose}>
+      <div
+        ref={panelRef}
+        className="popup-panel ask-evidence-layer"
+        role="dialog"
+        aria-modal="true"
+        aria-labelledby={headingId}
+        tabIndex={-1}
+        onClick={(event) => event.stopPropagation()}
+      >
+        <PopupCloseButton onClose={onClose} label={t("Close evidence panel")} />
+        <h2 id={headingId}>{postTitle}</h2>
+        {facts.length === 0 && images.length === 0 ? (
+          <p className="popup-placeholder">{t("No persisted evidence is available for this citation.")}</p>
+        ) : null}
+        {facts.length > 0 ? (
+          <section className="popup-section">
+            <h3>{t("Evidence facts")}</h3>
+            <ul className="post-evidence-list" aria-label={t("Evidence facts")}>
+              {facts.map((fact, index) => (
+                <li key={`${fact.kind}:${fact.text}:${index}`}>
+                  <span>{chatEvidenceKindLabel(fact.kind)}</span>
+                  <span>{fact.text}</span>
+                </li>
+              ))}
+            </ul>
+          </section>
+        ) : null}
+        {images.length > 0 ? (
+          <section className="popup-section">
+            <h3>{t("Image evidence")}</h3>
+            <ul className="post-evidence-list" aria-label={t("Image evidence")}>
+              {images.map((image) => (
+                <li key={image.unit_index}>
+                  <span>{image.caption ?? t("Untitled image")}</span>
+                  {image.extracted_text ? <span>{image.extracted_text}</span> : null}
+                </li>
+              ))}
+            </ul>
+          </section>
+        ) : null}
+        <button type="button" className="keyman-select" onClick={() => onOpenPost(postId)}>
+          {tf("Open post: {label}", { label: postTitle })}
+        </button>
+      </div>
+    </div>
+  );
+}
diff --git a/frontend/src/evidenceKindLabels.ts b/frontend/src/evidenceKindLabels.ts
new file mode 100644
index 000000000..ab92dabd4
--- /dev/null
+++ b/frontend/src/evidenceKindLabels.ts
@@ -0,0 +1,12 @@
+import { t } from "./i18n";
+
+const CHAT_EVIDENCE_KIND_LABELS: Record<string, string> = {
+  source_field: "Source field hint",
+  semantic_project: "Semantic project",
+  semantic_role: "Semantic role",
+  semantic_keyman: "Semantic Keyman",
+};
+
+export function chatEvidenceKindLabel(kind: string): string {
+  return t(CHAT_EVIDENCE_KIND_LABELS[kind] ?? "Evidence");
+}
diff --git a/frontend/src/i18n.ts b/frontend/src/i18n.ts
index 45a4e560b..482bbcb3f 100644
--- a/frontend/src/i18n.ts
+++ b/frontend/src/i18n.ts
@@ -167,8 +167,10 @@ const TRANSLATIONS: Partial<Record<Locale, Record<string, string>>> = {
     "Search related posts": "관련 글 검색",
     "Search related posts for: {name}": "{name} 관련 글 검색",
     "Evidence facts": "근거 사실",
+    "View evidence": "근거 보기",
     "Image evidence": "이미지 근거",
     "Untitled image": "제목 없는 이미지",
+    "No persisted evidence is available for this citation.": "이 인용에 사용할 수 있는 저장된 근거가 없습니다.",
     "Source field hint": "원천 필드 힌트",
     "Semantic project": "의미 기반 프로젝트",
     "Semantic role": "의미 기반 역할",
@@ -508,8 +510,10 @@ const TRANSLATIONS: Partial<Record<Locale, Record<string, string>>> = {
     "Search related posts": "搜索相关文章",
     "Search related posts for: {name}": "搜索与{name}相关的文章",
     "Evidence facts": "证据事实",
+    "View evidence": "查看证据",
     "Image evidence": "图像证据",
     "Untitled image": "无标题图像",
+    "No persisted evidence is available for this citation.": "此引用没有可用的已保存证据。",
     "Source field hint": "来源字段提示",
     "Semantic project": "语义项目",
     "Semantic role": "语义角色",
@@ -872,8 +876,10 @@ const TRANSLATIONS: Partial<Record<Locale, Record<string, string>>> = {
     "Search related posts": "関連投稿を検索",
     "Search related posts for: {name}": "{name}の関連投稿を検索",
     "Evidence facts": "証拠の事実",
+    "View evidence": "証拠を見る",
     "Image evidence": "画像証拠",
     "Untitled image": "無題の画像",
+    "No persisted evidence is available for this citation.": "この引用に使用できる保存済みの証拠はありません。",
     "Source field hint": "原典フィールドのヒント",
     "Semantic project": "意味的なプロジェクト",
     "Semantic role": "意味的な役割",
@@ -1212,8 +1218,10 @@ const TRANSLATIONS: Partial<Record<Locale, Record<string, string>>> = {
     "Search related posts": "Tìm bài viết liên quan",
     "Search related posts for: {name}": "Tìm bài viết liên quan đến {name}",
     "Evidence facts": "Sự kiện bằng chứng",
+    "View evidence": "Xem bằng chứng",
     "Image evidence": "Bằng chứng hình ảnh",
     "Untitled image": "Hình ảnh chưa đặt tên",
+    "No persisted evidence is available for this citation.": "Không có bằng chứng đã lưu cho trích dẫn này.",
     "Source field hint": "Gợi ý trường nguồn",
     "Semantic project": "Dự án ngữ nghĩa",
     "Semantic role": "Vai trò ngữ nghĩa",

From 47185cb0b5895620ca7b156a99fbcc7a93d4de52 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:01:16 -0700
Subject: [PATCH 02/11] test(ask): cover evidence dialog edge cases

---
 .../components/AskEvidenceLayerPopup.test.tsx | 32 +++++++++++++++++++
 1 file changed, 32 insertions(+)

diff --git a/frontend/src/components/AskEvidenceLayerPopup.test.tsx b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
index 314a8ac92..dac6a5147 100644
--- a/frontend/src/components/AskEvidenceLayerPopup.test.tsx
+++ b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
@@ -30,6 +30,9 @@ describe("AskEvidenceLayerPopup", () => {
     expect(screen.getByText(/project: Checkout revamp/)).toBeInTheDocument();
     expect(screen.getByText("Screenshot of the checkout error")).toBeInTheDocument();
     expect(screen.getByText("Error code 500 on checkout")).toBeInTheDocument();
+    expect(
+      screen.getByRole("list", { name: "Checkout error follow-up Evidence facts" }),
+    ).toBeInTheDocument();
   });
 
   it("shows an explicit placeholder when the citation has no persisted evidence", () => {
@@ -41,6 +44,19 @@ describe("AskEvidenceLayerPopup", () => {
     ).toBeInTheDocument();
   });
 
+  it("uses the untitled fallback when an image caption is blank", () => {
+    render(
+      <AskEvidenceLayerPopup
+        {...baseProps}
+        facts={[]}
+        images={[{ unit_index: 0, caption: "", extracted_text: null }]}
+        onClose={vi.fn()}
+        onOpenPost={vi.fn()}
+      />,
+    );
+    expect(screen.getByText("Untitled image")).toBeInTheDocument();
+  });
+
   it("closes on backdrop click, close button click, and Escape, but not on panel click", async () => {
     const onClose = vi.fn();
     const { container } = render(
@@ -82,4 +98,20 @@ describe("AskEvidenceLayerPopup", () => {
     );
     expect(screen.getByRole("dialog")).toHaveFocus();
   });
+
+  it("contains Tab and Shift+Tab focus within the modal layer", async () => {
+    render(
+      <AskEvidenceLayerPopup {...baseProps} facts={[]} images={[]} onClose={vi.fn()} onOpenPost={vi.fn()} />,
+    );
+    const closeButton = screen.getByRole("button", { name: "Close evidence panel" });
+    const openPostButton = screen.getByRole("button", { name: "Open post: Checkout error follow-up" });
+
+    openPostButton.focus();
+    await userEvent.tab();
+    expect(closeButton).toHaveFocus();
+
+    closeButton.focus();
+    await userEvent.tab({ shift: true });
+    expect(openPostButton).toHaveFocus();
+  });
 });

From ef302ba247c265d127f8cfcebbe683498cf115ae Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:01:51 -0700
Subject: [PATCH 03/11] fix(ask): contain evidence dialog focus

---
 .../src/components/AskEvidenceLayerPopup.tsx  | 55 +++++++++++++++++--
 1 file changed, 49 insertions(+), 6 deletions(-)

diff --git a/frontend/src/components/AskEvidenceLayerPopup.tsx b/frontend/src/components/AskEvidenceLayerPopup.tsx
index 1b9ca8b76..fe3c68c43 100644
--- a/frontend/src/components/AskEvidenceLayerPopup.tsx
+++ b/frontend/src/components/AskEvidenceLayerPopup.tsx
@@ -3,6 +3,15 @@ import { chatEvidenceKindLabel } from "../evidenceKindLabels";
 import { t, tf } from "../i18n";
 import { PopupCloseButton } from "./PopupCloseButton";
 
+const FOCUSABLE_SELECTOR = [
+  "a[href]",
+  "button:not([disabled])",
+  "input:not([disabled])",
+  "select:not([disabled])",
+  "textarea:not([disabled])",
+  '[tabindex]:not([tabindex="-1"])',
+].join(",");
+
 export type AskEvidenceLayerFact = {
   kind: string;
   text: string;
@@ -39,6 +48,8 @@ export function AskEvidenceLayerPopup({
   onOpenPost,
 }: AskEvidenceLayerPopupProps) {
   const headingId = useId();
+  const factsHeadingId = useId();
+  const imagesHeadingId = useId();
   const panelRef = useRef<HTMLDivElement>(null);
 
   useEffect(() => {
@@ -47,7 +58,39 @@ export function AskEvidenceLayerPopup({
 
   useEffect(() => {
     function handleKeyDown(event: KeyboardEvent) {
-      if (event.key === "Escape") onClose();
+      if (event.key === "Escape") {
+        event.preventDefault();
+        onClose();
+        return;
+      }
+      if (event.key !== "Tab") return;
+
+      const panel = panelRef.current;
+      if (!panel) return;
+      const focusable = Array.from(panel.querySelectorAll<HTMLElement>(FOCUSABLE_SELECTOR)).filter(
+        (element) => !element.hasAttribute("hidden") && element.getAttribute("aria-hidden") !== "true",
+      );
+      if (focusable.length === 0) {
+        event.preventDefault();
+        panel.focus();
+        return;
+      }
+
+      const first = focusable[0];
+      const last = focusable[focusable.length - 1];
+      const active = document.activeElement;
+      if (active === panel || !panel.contains(active)) {
+        event.preventDefault();
+        (event.shiftKey ? last : first).focus();
+        return;
+      }
+      if (event.shiftKey && active === first) {
+        event.preventDefault();
+        last.focus();
+      } else if (!event.shiftKey && active === last) {
+        event.preventDefault();
+        first.focus();
+      }
     }
     document.addEventListener("keydown", handleKeyDown);
     return () => document.removeEventListener("keydown", handleKeyDown);
@@ -71,8 +114,8 @@ export function AskEvidenceLayerPopup({
         ) : null}
         {facts.length > 0 ? (
           <section className="popup-section">
-            <h3>{t("Evidence facts")}</h3>
-            <ul className="post-evidence-list" aria-label={t("Evidence facts")}>
+            <h3 id={factsHeadingId}>{t("Evidence facts")}</h3>
+            <ul className="post-evidence-list" aria-labelledby={`${headingId} ${factsHeadingId}`}>
               {facts.map((fact, index) => (
                 <li key={`${fact.kind}:${fact.text}:${index}`}>
                   <span>{chatEvidenceKindLabel(fact.kind)}</span>
@@ -84,11 +127,11 @@ export function AskEvidenceLayerPopup({
         ) : null}
         {images.length > 0 ? (
           <section className="popup-section">
-            <h3>{t("Image evidence")}</h3>
-            <ul className="post-evidence-list" aria-label={t("Image evidence")}>
+            <h3 id={imagesHeadingId}>{t("Image evidence")}</h3>
+            <ul className="post-evidence-list" aria-labelledby={`${headingId} ${imagesHeadingId}`}>
               {images.map((image) => (
                 <li key={image.unit_index}>
-                  <span>{image.caption ?? t("Untitled image")}</span>
+                  <span>{image.caption?.trim() ? image.caption : t("Untitled image")}</span>
                   {image.extracted_text ? <span>{image.extracted_text}</span> : null}
                 </li>
               ))}

From 9f5b0a0bb290af19e54af5376ee69c28ed2b0e69 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:02:12 -0700
Subject: [PATCH 04/11] docs(storybook): cover blank evidence caption

---
 .../components/AskEvidenceLayerPopup.stories.tsx  | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff --git a/frontend/src/components/AskEvidenceLayerPopup.stories.tsx b/frontend/src/components/AskEvidenceLayerPopup.stories.tsx
index f542f50d1..070ba40f5 100644
--- a/frontend/src/components/AskEvidenceLayerPopup.stories.tsx
+++ b/frontend/src/components/AskEvidenceLayerPopup.stories.tsx
@@ -77,3 +77,18 @@ export const UntitledImage: Story = {
     ],
   },
 };
+
+// Edge case: some sources persist an empty caption rather than null. The buyer
+// still needs a visible label that explains what to do with the evidence row.
+export const BlankImageCaption: Story = {
+  args: {
+    facts: [],
+    images: [
+      {
+        unit_index: 0,
+        caption: "",
+        extracted_text: "Diagram OCR text remains available",
+      },
+    ],
+  },
+};

From 619ea7b66e079e1fe8cd4d2581476673b5031245 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:03:04 -0700
Subject: [PATCH 05/11] docs(ask): trace modal accessibility standard

---
 docs/doctoring/DESIGN_TOKEN_REFERENCES.md | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/docs/doctoring/DESIGN_TOKEN_REFERENCES.md b/docs/doctoring/DESIGN_TOKEN_REFERENCES.md
index 717954045..c22317628 100644
--- a/docs/doctoring/DESIGN_TOKEN_REFERENCES.md
+++ b/docs/doctoring/DESIGN_TOKEN_REFERENCES.md
@@ -1,8 +1,8 @@
 # Design-token and Storybook traceability
 
 **Status:** Active PR evidence; not protected-main truth until merge.  
-**Scope:** `frontend/src/styles/tokens.css`, repeated chip/close modules, and
-the Storybook inventory.
+**Scope:** `frontend/src/styles/tokens.css`, repeated chip/close modules, the
+Ask evidence dialog, and the Storybook inventory.
 
 ## Standards mapped to implementation
 
@@ -11,6 +11,7 @@ the Storybook inventory.
 | W3C Design Tokens Format Module 2025.10 | Name color, space, type, and radius once; consume those names from repeated objects. | `frontend/src/styles/tokens.css` defines `--color-*`, `--space-*`, `--size-control-min`, `--radius-chip`, `--radius-control`, `--radius-panel`, and `--font-*`. `CitationChip`, `PopupCloseButton`, `CutoffKnownBody`, and `LineageEntityPicker` read those names through `App.css`. |
 | Storybook for React & Vite | Catalog repeated controls so a buyer can try the next click without reading `App.tsx`. | `frontend/src/components/*.stories.tsx` and `docs/storybook-inventory.md`. |
 | WCAG 2.2 | Give interactive controls programmatic names and announce an asynchronous evidence failure instead of leaving a perpetual loading state. | Component interaction tests exercise the named controls; `EvidencePanel` exposes its terminal failure with `role="alert"`. This is targeted evidence, not a claim of complete WCAG conformance. |
+| WAI-ARIA APG Dialog (Modal) Pattern | A surface marked `aria-modal="true"` must behave modally: focus moves inside, `Tab` and `Shift+Tab` remain inside, and `Escape` closes the layer. | `AskEvidenceLayerPopup` moves initial focus inside the dialog and explicitly cycles forward/backward keyboard focus between its actionable controls; component tests cover both focus-loop directions and Escape. Its evidence lists use dialog-specific accessible labels so assistive technology can distinguish the modal list from the still-rendered inline answer. |
 
 ## APA 7th references
 
@@ -23,3 +24,7 @@ https://storybook.js.org/docs/get-started/frameworks/react-vite
 
 World Wide Web Consortium. (2024). *Web Content Accessibility Guidelines
 (WCAG) 2.2* (W3C Recommendation). https://www.w3.org/TR/WCAG22/
+
+World Wide Web Consortium. (n.d.). *Dialog (modal) pattern*. WAI-ARIA
+Authoring Practices Guide. Retrieved August 22, 2026, from
+https://www.w3.org/WAI/ARIA/apg/patterns/dialog-modal/

From 51ae72f97eb52603f8acd87a730c935a252e5e4e Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:03:23 -0700
Subject: [PATCH 06/11] docs(storybook): inventory Ask evidence layer

---
 docs/storybook-inventory.md | 1 +
 1 file changed, 1 insertion(+)

diff --git a/docs/storybook-inventory.md b/docs/storybook-inventory.md
index 28c59bd48..a92dccf02 100644
--- a/docs/storybook-inventory.md
+++ b/docs/storybook-inventory.md
@@ -6,6 +6,7 @@ buyer-facing control you can click before changing product CSS.
 | Story | Buyer next action | Token / module |
 |---|---|---|
 | `Evidence/CitationChip` | Click a cited title to open that source post. | `--color-chip-border`, `--radius-chip`, `CitationChip` |
+| `Evidence/AskEvidenceLayerPopup` | Inspect one citation without leaving the answer; close to continue the answer or open the complete source post. Stories cover text/image evidence, no-evidence, missing OCR, null caption, and blank-caption fallback states. | shared popup tokens through `App.css`, `PopupCloseButton`, `AskEvidenceLayerPopup` |
 | `AnalysisRun/CutoffKnownBody` | Read the cutoff-known sentence, then compare it with the live body below. | `--color-accent-border`, `--space-panel-block`, `--radius-panel`, `CutoffKnownBody` |
 | `Analysis/LineageEntityPicker` | Choose which corp to reconstruct, then click Request a lineage reconstruction. | `--space-control-gap`, `--size-control-min`, `--radius-control`, `LineageEntityPicker` |
 | `Chrome/PopupCloseButton` | Close the evidence panel or post popup. | `--space-close-inset`, `--font-size-close`, `PopupCloseButton` |

From 7a9481ea279475aca96e9dc429be55ac636329c4 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:03:50 -0700
Subject: [PATCH 07/11] docs(changelog): record Ask evidence layer

---
 CHANGELOG.md | 2042 +-------------------------------------------------
 1 file changed, 12 insertions(+), 2030 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index c8ed1a099..ae29adc82 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,8 +6,18 @@ All notable changes to this project are documented here. Format follows
 
 ## [Unreleased]
 
+### Added
+
+- Ask Agent citations can now open a focused evidence layer that preserves the
+  current answer while exposing persisted text and image evidence, with
+  Storybook coverage for populated, empty, missing-OCR, null-caption, and
+  blank-caption states.
+
 ### Fixed
 
+- The Ask citation evidence layer now contains forward and reverse keyboard
+  focus while marked modal, gives its modal evidence lists citation-specific
+  accessible names, and falls back to `Untitled image` for blank captions.
 - `make smoke` and `make seed` now run through the locked project `uv`
   environment, so local OIDC and synthetic-data workflows resolve the same
   pinned dependencies as CI.
@@ -71,2033 +81,5 @@ All notable changes to this project are documented here. Format follows
   constraint). Root cause not yet conclusively identified; deferred
   as out of scope for this migration-catchup change (a different
   feature area -- analysis-run/TEPP lifecycle, not R&R/summary/
-  verification) rather than rushed. 553 other tests unaffected.
-
-- `get_or_create_corporate_entity`'s post-lock duplicate-create re-check
-  fuzzy-matched against every cataloged entity, not just an exact
-  concurrent duplicate of the entity being created. A newly-created
-  parent whose name is a prefix of the child now being created (e.g.
-  "Acme" as parent of "Acme Gwangju Plant") scored ~0.7 similarity
-  against that child under the shared 0.6 threshold, so the child was
-  silently bound to its own parent's id instead of getting its own
-  catalog row -- undermining exactly the "통합 고객사 계열 tree AI"
-  (integrated customer affiliate tree) hierarchy the feature exists
-  for. The re-check now requires an exact post-normalization match
-  (`min_similarity=1.0`); real mention resolution against the full
-  candidate set is unchanged. Caught locally by
-  `test_first_mention_of_a_new_counterparty_creates_a_real_corporate_entity`,
-  which requires a live PostgreSQL/Keycloak/Valkey stack and is
-  therefore skipped in CI (`make up` required) -- confirmed CI's own
-  "Full test suite" run has never actually executed this assertion.
-- `test_start_analysis_run_recovers_the_a100_fork` seeded
-  `snapshot_sha256`/`configuration_sha256`/`code_revision_sha` with
-  `"t"`/`"u"`/`"v"`-repeated literals; none are valid hex characters,
-  so the very first insert failed its own `analysis_source_snapshot`
-  check constraint every time this test actually ran. Same CI-blind
-  gap as above -- fixed to valid hex placeholders matching this file's
-  existing convention.
-
-## [2.12.3] - 2026-08-18
-
-### Added
-
-- `make seed` inserts Late Demo public post (2026-01-13) so the
-  January 12 Demo Corp lineage and TEPP runs' knowledge cutoff is
-  falsifiable: Demo public post still opens; Late Demo does not.
-  The live post list still shows Late Demo. The cutoff filter
-  itself already lives on this stack (ADR 0016). TEPP honesty is
-  unchanged: accepted acks stay Failed transport evidence, not
-  Succeeded. Never invent a theta.
-
-## [2.12.2] - 2026-08-18
-
-### Fixed
-
-- Accepted TEPP transport evidence now stores **received** (transport
-  response) and **recorded** (row write) as distinct clocks when those
-  instants differ (ADR 0035 follow-up). After `make seed`, Demo Analyst
-  opens **TEPP measurement · Failed · Demo Corp** Measurement evidence
-  and sees one Received clock when seed receipt and persist share an
-  instant. A later start that persists in a later minute shows both
-  clocks. Digest recomputation is unchanged. Hidden runs stay 404.
-  Never invent a theta.
-
-## [2.12.1] - 2026-08-17
-
-### Fixed
-
-- A published TEPP **accepted** acknowledgement is stored as
-  **aggregate transport evidence** and stays Failed /
-  `tepp_completed_result_unsupported` (ADR 0035). After `make seed`,
-  Demo Analyst opens that Failed Demo Corp row to read contract
-  version, accepted run id, clocks, and a copyable SHA-256. The
-  section says completed-artifact identity is unavailable until TEPP
-  publishes a versioned completed-result contract. A
-  LineageWeave-local `time_multilevel_multi_affiliation` envelope, or
-  any other unpublished completed shape, stays Failed /
-  `tepp_result_not_persisted` and must not stamp Succeeded. Missing
-  `TEPP_TRANSPORT_URL` stays Failed / `tepp_not_available`. Never
-  invent a theta.
-
-## [2.12.0] - 2026-08-17
-
-### Added
-
-- A persistable TEPP **time / multilevel / multi-affiliation** result
-  is stored on the analysis-run and marked Succeeded (ADR 0034). After
-  `make seed`, Demo Analyst sees **TEPP measurement · Succeeded · Demo
-  Corp** next to the Failed missing-transport row. Home list and detail
-  show measured clocks and affiliation counts. A screen reader on that
-  Succeeded row hears open the run to read those aggregates, not only
-  the title. An `accepted` ack or an envelope this product cannot store
-  stays Failed / `tepp_result_not_persisted`. Missing
-  `TEPP_TRANSPORT_URL` stays Failed / `tepp_not_available`. Never
-  invent a theta.
-
-## [2.11.0] - 2026-08-17
-
-### Added
-
-- Home now shows the authorized customer-group tree (Group / Company /
-  Plant) instead of only a flat corp list. After `make seed`, Demo
-  Analyst walks Demo Group → Demo Corp → Demo Plant; a click opens that
-  entity as the corporate-entity report grouping. The post-scoped
-  affiliate tree is unchanged (ADR 0033).
-- Abbreviations on a post are cross-checked against that tree through
-  the existing Searxng client. A unique corroborated hit binds; a down,
-  empty, or tied search stays unbound and does not invent a parent or
-  AUTO row. Seeded `DC` on Demo Corp is synthetic Demo Corp only.
-
-## [2.10.4] - 2026-08-17
-
-### Fixed
-
-- After a Demo Corp lineage reconstruction has started, the same
-  granted retention purge empties `analysis_run_lineage_edge`,
-  `analysis_run_reconstruction`, and `analysis_source_snapshot_member`
-  when those tables exist, including their delete-reject triggers
-  (ADR 0032). Follow the same grant + admin + phrase path — do not
-  `DISABLE TRIGGER` as superuser.
-
-## [2.10.3] - 2026-08-17
-
-### Fixed
-
-- Opening a listed analysis-run that then 404s drops that stale row
-  from the home list after an authorized re-read. The next action is
-  announced as a status alert: open a remaining visible run, or
-  request a lineage reconstruction. The message still does not name
-  the thread or the cutoff (ADR 0014 / ADR 0018).
-
-## [2.10.2] - 2026-08-17
-
-### Fixed
-
-- Opening a post whose embedded picture uses invoice-like HTML
-  (`alt="Invoice > 1000"`, unquoted `width`, newlines in the base64)
-  now shows the picture. The raw payload no longer returns when a
-  remote-only or SVG tag is the whole body. Re-export as PNG or JPEG
-  if the type is rejected. The popup, `extract_base64_images`, and
-  `chunk_by_dom` share one raster allowlist (ADR 0031).
-
-## [2.10.1] - 2026-08-17
-
-### Fixed
-
-- Analysis-run list buttons now include the kind-specific next-action
-  sentence in the accessible name (WCAG 2.2 SC 4.1.2). Open a Failed
-  TEPP row: a screen reader hears connect the measurement service, not
-  only the run title. `aria-label` replaces button contents (ADR 0014).
-  No TEPP theta is invented.
-
-## [2.10.0] - 2026-08-17
-
-### Added
-
-- Home Rankings panel fuses visible posts through `RankWeaveClient`
-  (ADR 0030). After login with the port disabled or the library
-  missing, Demo Analyst sees **Rankings · RankWeave not available**.
-  An accepted hit lists the title; click opens that post. A hidden
-  post is omitted. Never invent a fused score or a theta.
-- Period reports now persist closest and farthest leftover
-  post–criterion pairs after the IRT main effects (Jeon leftover
-  map, ADR 0028 / 0029). After `make seed`, leftover pairs sit above
-  the member list; clicking a pair opens that post. A leftover pair
-  for a hidden post is omitted the same way a hidden member is.
-
-## [2.9.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  next action after landed cited evidence: Linked post evidence is
-  current, then read Event Lineage on that post. Home list opens do
-  not add that copy. No TEPP theta is invented. No cutoff body is
-  invented (ADR 0016).
-
-## [2.8.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now puts the
-  first cited evidence immediately under the named citation next
-  action, ahead of the chat input. Home list opens still wait for a
-  citation click. No TEPP theta is invented. No cutoff body is
-  invented (ADR 0016).
-
-## [2.7.2] - 2026-08-17
-
-### Fixed
-
-- R&R person chips now read `cataloged_person_id` from
-  `post_summary_role` (ADR 0027). Open a post whose R&R names a
-  cataloged person: the chip is a button even when Keyman extraction
-  was not run on that post. Click it to walk that person, not a later
-  same-named row. Historical backfill leaves a role unbound when two
-  same-named mentions already exist.
-
-## [2.7.1] - 2026-08-17
-
-### Fixed
-
-- `POST /api/analysis-runs` records Pending lineage only (ADR 0017).
-  TEPP and period-report kinds are 422 so this path cannot invent a
-  measurement. Open Analysis runs and wait until affiliated corps
-  load; choose a corp if you walk more than one, then click
-  **Request a lineage reconstruction**. Preview the picker in
-  Storybook (`Analysis/LineageEntityPicker`). Failed TEPP stays
-  terminal on this write.
-
-## [2.7.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  first cited source after the landed first Ask answer: Linked post,
-  then open that evidence. Home list opens do not add that copy. No
-  TEPP theta is invented. No cutoff body is invented (ADR 0016).
-
-## [2.6.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now puts the
-  first Ask answer immediately under the named seed next action, ahead
-  of the chat input. Home list opens keep that answer after the input.
-  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
-
-## [2.5.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  first Ask after landed chat: What happened between these events,
-  then read that answer. Home list opens do not add that copy. No
-  TEPP theta is invented. No cutoff body is invented (ADR 0016).
-
-## [2.4.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now puts Ask
-  about this lineage immediately under the Ask next action, ahead of
-  Affiliate tree. Home list opens keep Ask after Keyman. No TEPP theta
-  is invented. No cutoff body is invented (ADR 0016).
-
-## [2.3.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now focuses
-  the Ask heading after Priya Nair related nodes are current. Home
-  list opens do not steal that focus. No TEPP theta is invented. No
-  cutoff body is invented (ADR 0016).
-
-## [2.2.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  next action after Priya Nair related nodes are current: **Ask about
-  this lineage**. Home list opens do not add that copy. No TEPP theta
-  is invented. No cutoff body is invented (ADR 0016).
-
-## [2.1.0] - 2026-08-17
-
-### Added
-
-- Opening a title marked **Updated after cutoff** now shows the body
-  that run knew beside the live rewrite. After `make seed`, open Demo
-  public post from the Demo Corp lineage run: **Body this run knew** is
-  the January follow-up; the live body names the later delivery window.
-  `GET /api/posts/{id}?as_of=` reads `source_post_revision`. Analysis-run
-  detail stays titles and clocks. A missing revision is omitted — never
-  a fabricated cutoff sentence or a TEPP theta (ADR 0025).
-
-## [2.0.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now lands
-  Priya Nair related nodes under the first-related next action, ahead
-  of Affiliate tree. Home list opens still wait for a related click.
-  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
-
-## [1.9.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  first related node after landed Ada West related: Priya Nair, then
-  read that person. Home list opens do not add that copy. No TEPP
-  theta is invented. No cutoff body is invented (ADR 0016).
-
-## [1.8.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now lands
-  Ada West related nodes under the first-Keyman next action, ahead
-  of Affiliate tree. Home list opens still wait for a Keyman click.
-  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
-
-## [1.7.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now names the
-  first Keyman after landed evaluation: Ada West, then read that
-  person. Home list opens do not add that copy. No TEPP theta is
-  invented. No cutoff body is invented (ADR 0016).
-
-## [1.6.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members now puts
-  Keyman and evaluation immediately under the Event Lineage next
-  action, ahead of Affiliate tree. Home list opens keep evaluation
-  above Event Lineage. No TEPP theta is invented. No cutoff body is
-  invented (ADR 0016).
-
-## [1.5.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members names the next
-  action after the current Event Lineage node: read Keyman and
-  evaluation. Home list opens do not add that copy. No TEPP theta is
-  invented. No cutoff body is invented (ADR 0016).
-
-## [1.4.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members marks that
-  post current in the popup Event Lineage DAG, so the focused heading
-  has a you-are-here node. The home DAG stays unmarked. No TEPP theta
-  is invented. No cutoff body is invented (ADR 0016).
-
-## [1.3.0] - 2026-08-17
-
-### Added
-
-- Opening Public post from the landed Demo Corp members focuses the
-  popup Event Lineage heading, matching the next-action copy. Home
-  post-list opens do not steal that focus. No TEPP theta is invented.
-
-## [1.2.0] - 2026-08-17
-
-### Added
-
-- Opening **Public post** from the landed Demo Corp report now names
-  the next action: that post is open from Demo Corp, so read Event
-  Lineage, Keyman, and evaluation. The opened member is current. Mean
-  θ stays on the report panel. No TEPP theta is invented. No cutoff
-  body is invented (ADR 0016).
-
-## [1.1.0] - 2026-08-17
-
-### Added
-
-- Opening **Open period report 2026-W02** now puts the Demo Corp
-  report (mean θ and member posts) immediately under the named next
-  action, ahead of Other Corp and the week strip. The Public post
-  member stays clickable. Mean θ stays on the report panel. No TEPP
-  theta is invented. No cutoff body is invented (ADR 0016).
-
-## [1.0.0] - 2026-08-17
-
-### Added
-
-- Opening **Open period report 2026-W02** now names the next action on
-  the landed Demo Corp report: read its mean θ and member posts, then
-  open a post. The focused comparison chip uses the visible
-  `Corporate entity: Demo Corp` caption and the persisted mean θ
-  (WCAG 2.5.3). Changing the week still focuses the report period
-  field. Mean θ stays on the report panel. No TEPP theta is invented.
-
-## [0.99.0] - 2026-08-17
-
-### Added
-
-- Opening **Open period report 2026-W02** when the operator is already
-  on that week lands the grouping comparison strip on Demo Corp. The
-  Demo Corp chip is current and focused. Changing the week still
-  focuses the report period field. Mean θ stays on the report panel.
-  No TEPP theta is invented.
-
-## [0.98.1] - 2026-08-17
-
-### Fixed
-
-- Authorized analysis-run list and detail queries are now complete
-  parameterized SQL literals. Semgrep no longer treats the visibility
-  predicate as string-concatenated user input. The $1 / $2 / $3 binds
-  are unchanged. No TEPP theta is invented.
-
-## [0.98.0] - 2026-08-17
-
-### Added
-
-- Opening **Open period report 2026-W02** from a corporate-entity
-  analysis run also switches Report grouping to Corporate entity and
-  marks the Demo Corp grouping current. The opened report is named
-  Demo Corp, not a UUID. Mean θ stays on the report panel. No TEPP
-  theta is invented.
-
-## [0.97.0] - 2026-08-17
-
-### Added
-
-- A Succeeded period-report analysis run now opens the scored week.
-  After `make seed`, open **Period report · Succeeded · Demo Corp**
-  and click **Open period report 2026-W02**: the report period field
-  is focused on that week. Failed rows stay closed. Mean θ stays on
-  the report panel. No TEPP theta is invented.
-
-## [0.96.0] - 2026-08-17
-
-### Added
-
-- `make seed` now records **Period report · Succeeded · Demo Corp** on
-  the shared snapshot after the calibrated report tables are written
-  (ADR 0024). Open that row to confirm the cutoff posts. Mean θ stays
-  on the period-report panel. Start stays 422. No TEPP theta is
-  invented.
-
-## [0.95.0] - 2026-08-17
-
-### Added
-
-- Analysis-run detail now lists labeled outbox delivery: Claimed then
-  Delivered (ADR 0023). After `make seed`, open the Demo Corp lineage
-  run to see those times. Stream entry ids stay off the payload. No
-  TEPP theta is invented.
-
-## [0.94.0] - 2026-08-17
-
-### Added
-
-- `POST /api/analysis-runs/{id}/start` now commits Running plus one
-  durable outbox row, wakes Valkey (`analysis-run-outbox`), then
-  delivers ThreadWeave or `tepp_client` (ADR 0023). A crash after
-  Start leaves the work item; refresh finishes it. Period-report
-  stays 422. No TEPP theta is invented.
-
-## [0.93.0] - 2026-08-17
-
-### Added
-
-- `make seed` now persists the designed A-100 fork on the Demo Corp
-  Succeeded lineage run. Open that run: the revised quote and delivery
-  question follow the pricing follow-up and are buttons. Start is
-  unchanged. No TEPP theta is invented.
-
-## [0.92.0] - 2026-08-17
-
-### Added
-
-- **Start TEPP measurement** on a Pending TEPP row submits
-  `AnalysisRunRequest` through `tepp_client` (ADR 0022). A missing
-  `TEPP_TRANSPORT_URL` or a refused URL is Failed /
-  `tepp_not_available`. An accepted envelope is Failed /
-  `tepp_result_not_persisted`. Failed stays terminal: **Request a new
-  TEPP measurement** records a new Pending run. Period-report start
-  stays 422. No TEPP theta is invented.
-
-## [0.91.0] - 2026-08-17
-
-### Added
-
-- After **Start reconstruction**, the titled A-100 edges are buttons.
-  Click the revised-quote child to open the live post; click the
-  pricing-follow-up parent to open that post. A child marked
-  **Updated after cutoff** still shows the live-body warning. The
-  popup does not invent a cutoff snapshot. No TEPP theta is invented.
-
-## [0.90.0] - 2026-08-17
-
-### Added
-
-- Opening an analysis-run title marked **Updated after cutoff** now
-  shows a popup status that the body is live, not a cutoff snapshot
-  (ADR 0016). After `make seed`, open the Demo Corp lineage run and
-  click Demo public post: the warning appears above the live body
-  and tells you to compare it with this run. Demo private post and
-  the home post list do not. The popup does not invent the earlier
-  text. No TEPP theta is invented.
-
-## [0.89.0] - 2026-08-17
-
-### Added
-
-- Analysis-run detail now compares each in-cutoff title's live
-  `updated_at` with that run's knowledge cutoff. After `make seed`,
-  open the Demo Corp lineage run: Demo public post is marked
-  **Updated after cutoff**; Demo private post is not. Opening a
-  marked title still shows the live body -- cutoff body versioning
-  stays a later slice (ADR 0016). The list stays aggregates-only.
-  No TEPP theta is invented.
-
-## [0.88.0] - 2026-08-16
-
-### Added
-
-- `POST /api/analysis-runs/{id}/start` runs ThreadWeave on a visible
-  Pending lineage cutoff bag and persists run-scoped parent choices
-  (ADR 0021). Open the Pending Demo Corp row, then start reconstruction.
-  The designed A-100 fork (revised quote and delivery question under the
-  pricing follow-up) is the acceptance tree. TEPP and period-report
-  start are 422 — this path does not invent a theta. A Succeeded retry
-  returns the stored digest. A Running restart is 409. Create freezes
-  authorized post ids so start cannot pick up a later backfill. Live
-  Event Lineage stays a separate rebuild.
-
-## [0.87.0] - 2026-08-16
-
-### Added
-
-- Operators can empty a run-bearing analysis-run registry without a
-  superuser trigger disable. Insert an unrevoked
-  `analysis_run_retention_grant` for `session_user`, grant
-  `analysis_run_retention_admin`, then
-  `select purge_analysis_run_registry('approved-retention-purge')`
-  (ADR 0020). Export `analysis_run_retention_event`, delete those
-  rows, then roll back 0020 and 0018. A raw `DELETE`, a published
-  token without a grant, and a runtime role that is not the admin
-  role still fail.
-- Repeated citation chips and close buttons use named design tokens
-  in `frontend/src/styles/tokens.css`. Preview them in Storybook
-  (`cd frontend && pnpm run storybook`).
-
-## [0.86.2] - 2026-08-16
-
-### Fixed
-
-- An R&R organization button now walks the catalog id stored on that
-  role row (ADR 0019). Two catalog orgs can share a display name; open
-  the post, click the name, and you stay on the resolved org — not a
-  homonym. `GET /api/teams/{id}/related` matches person/entity authz:
-  another corp's private-only team is 403; an unknown UUID is 404.
-
-## [0.86.1] - 2026-08-16
-
-### Changed
-
-- Opening a post or its evidence panel now shows each embedded
-  `data:image` picture in document order, with the surrounding sentences
-  as text. The raw base64 string is no longer dumped into the popup.
-  Remote `http(s)` image URLs stay unloaded. After `make seed`, a post
-  whose body includes a data-URI image shows the picture; Extract Keyman
-  or Ask still runs OCR on that image for search.
-
-### Fixed
-
-- Opening a Pending lineage run repeats that reconstruction has not
-  started. Pending next-action copy is pinned to the registered run
-  kinds, so a Pending TEPP row does not say reconstruction.
-
-## [0.86.0] - 2026-08-16
-
-### Added
-
-- Related-node walks now include team and organization mention edges.
-  After `make seed` and a summary that names 설계팀 on two posts, open
-  either post, click the R&R team, and open the sibling post (ADR 0018).
-  A team-only follow-up is no longer an island.
-- `GET /api/teams/{team_id}/related` starts the same RWR walk Keyman
-  and corporate-entity related already use. Related team chips are
-  buttons.
-
-### Fixed
-
-- Thread-group analysis-run *lists* now require an in-cutoff visible
-  post. A later public post in that thread group no longer surfaces a
-  January run the account was not allowed to know.
-- Failed period-report rows tell the operator to rebuild the report.
-  Next-action copy is pinned to the registered run kinds. A pending
-  TEPP corpus does not claim a calibrated measurement.
-
-## [0.85.0] - 2026-08-16
-
-### Added
-
-- `POST /api/analysis-runs` records a Pending lineage or TEPP run on an
-  authorized cutoff capture (ADR 0017). The home panel's **Request a
-  lineage reconstruction** button writes that row so an operator can
-  confirm the cutoff corpus immediately. Reconstruction and live TEPP
-  execution stay later slices — this write never invents a theta.
-- Failed lineage rows tell the operator to retry reconstruction; only
-  Failed TEPP rows mention the measurement service. Pending rows say
-  reconstruction has not started yet.
-
-## [0.84.1] - 2026-08-16
-
-### Fixed
-
-- Analysis-run detail keeps 12-character digest prefixes as visible
-  text (so assistive technology hears `Code` / `Config` values) and
-  puts the full digest on hover. Open the Demo Corp lineage run, hover
-  a prefix, and match it to the API payload. The home list still hides
-  digests even when the list JSON includes them.
-- Opening a cutoff title now says the live body may have changed after
-  that run. Compare the opened post with the cutoff date before you
-  treat it as reconstructed evidence (ADR 0016).
-
-## [0.84.0] - 2026-08-16
-
-### Added
-
-- `make seed` records a Demo Corp TEPP measurement run through
-  `tepp_client` on the same snapshot as the lineage run (ADR 0013).
-  The default transport is unavailable, so the home list shows
-  "TEPP measurement · Failed · Demo Corp" and tells the operator to
-  open the run, then connect the measurement service. Detail history
-  keeps `tepp_not_available` -- never a fabricated theta. TEPP stays
-  a wire client, not a local psychometric engine. `make seed` skips
-  snapshot-count inserts once counts exist so a re-run does not hit
-  the freeze trigger. A failed lineage row tells the operator to retry
-  reconstruction; only a failed TEPP row mentions the measurement
-  service. A failed period-report row tells the operator to rebuild
-  the report from a current snapshot. A pending TEPP row does not
-  claim a calibrated measurement. Stacked PRs now run the same
-  GitHub Checks as PRs to main.
-
-## [0.83.0] - 2026-08-16
-
-### Fixed
-
-- Analysis-run detail now lists only ABAC-visible posts whose
-  `created_at` is at or before that run's `knowledge_cutoff`. After
-  `make seed`, open the Demo Corp lineage run: Demo public post is
-  there; a later own-corp follow-up is not. The live post list is
-  unchanged. Click a listed title to inspect what that cutoff
-  reconstructed (ADR 0016).
-- Upgrading through `0016_cross_post_actor_identity.sql` copies R&R
-  person names into `post_summary_person_mention` and leaves Keyman
-  `post_person_mention.mention_context` in place. Re-run Keyman only
-  when you want a new Keyman set -- a later summary no longer erases
-  the stolen row.
-
-## [0.82.0] - 2026-08-16
-
-### Added
-
-- Analysis-run detail lists ABAC-visible posts in the run's scope.
-  After `make seed`, the Demo Corp lineage run opens the Demo public
-  post. Hidden other-corp private posts never appear. List payloads
-  stay aggregates-only.
-
-## [0.81.0] - 2026-08-16
-
-### Added
-
-- Analysis-run detail shows the labeled lifecycle: Pending, Running,
-  then Succeeded, with occurrence times from `analysis_run_status_event`.
-  The list stays latest-status only. Hidden runs still 404 and never
-  leak events. Failure codes stay machine tokens -- no invented label.
-  Synthetic Demo Corp seed only.
-
-## [0.80.0] - 2026-08-16
-
-### Added
-
-- Home Analysis runs rows are buttons. Clicking the seeded Demo Corp
-  lineage run opens `GET /api/analysis-runs/{id}` and shows cutoff,
-  requested date, and document count. A hidden run is "This analysis
-  run is not visible." -- never a raw 404 or a DSN. Still synthetic
-  aggregates only.
-
-## [0.79.0] - 2026-08-16
-
-### Added
-
-- Authorized analysis-run evidence on the product home page. After
-  `make seed`, Demo Analyst sees "Lineage reconstruction · Succeeded ·
-  Demo Corp" with the synthetic document count. `GET /api/analysis-runs`
-  is scoped in SQL: another tenant's run 404s and never appears in the
-  list. The payload is labels and aggregates -- never source SQL, a DSN,
-  or a raw record. TEPP stays behind `tepp_client`; Null channels are
-  unchanged.
-
-## [0.78.0] - 2026-08-15
-
-### Changed
-
-- Related-node person chips now use the localized `person_side` lookup label
-  supplied by the authorized API payload. Users see business context such as
-  `Our side` or `Counterparty`, while ontology class metadata remains available
-  separately for semantic processing and provenance.
-- Related-person buttons now expose that same caption in the accessible name, so
-  assistive technology hears `Related nodes for Priya Nair (Counterparty)`
-  instead of the name alone.
-- Structured extraction, summarization, commitment, relationship-classification,
-  and LLM-as-a-Judge consumers now request contextual-orchestrator `auto` mode
-  so the orchestration plane can meet the quality requirement and then minimize
-  known execution cost. Explicit checked `verify` paths remain unchanged
-  (ADR 0015).
-
-## [0.77.0] - 2026-08-14
-
-### Fixed
-
-- Keyman and R&R person mentions now replace independent source projections. Knowledge Graph edges have one canonical identity plus post-level evidence, so removed actors and concurrent writes cannot leave stale or duplicate buyer-visible relationships.
-- Vision-response parsing now strips balanced outer Markdown emphasis from field values
-  while still accepting emphasized field labels, so OCR such as
-  ``TEXT: **LT7**`` is not truncated.
-- A real live synthetic regression batch run surfaced a genuine
-  `DeadlockDetectedError` from concurrent corporate-entity creation:
-  two concurrent transactions each creating a different new entity,
-  mentioned in opposite order across two different posts, took
-  row-level locks in opposite order and deadlocked. Entity *creation*
-  (the rare, first-mention-only branch) now serializes through a
-  single named Postgres advisory transaction lock, taken only right
-  before the write and auto-released at commit/rollback -- the
-  lock-free similarity-matching fast path every already-cataloged
-  entity resolves through is unaffected. See ADR 0012.
-
-## [0.76.0] - 2026-08-14
-
-### Added
-
-- Standards-complete W3C PROV-O support: all 30 classes, all 50
-  normative properties, both qualification tables, qualified-to-
-  unqualified implication, property hierarchy, defined inverses,
-  Appendix B inverse-name normalization, RDF serialization, and a
-  normalized PostgreSQL assertion store with fail-closed domain,
-  range, object-kind, and datatype enforcement (ADR 0011).
-- A dedicated exact-head PROV-O contract workflow runs the complete
-  registry/inference suite, real PostgreSQL migration tests, public
-  docstring checks, and 100% statement/branch coverage for the owned
-  runtime module.
-
-### Changed
-
-- The product navigation graph remains an explicit projection;
-  literal-valued and qualified provenance is no longer forced into
-  `knowledge_graph_edge`.
-
-### Fixed
-
-- Review hardening verifies complete hierarchy placement, rejects parent
-  failures and cycles, propagates canonical affiliations, replaces stale
-  actor projections, enforces atomic team identity, validates timezone-aware
-  `xsd:dateTime` literals (including the XSD `±14:00` offset bound), and
-  protects referenced provenance rows.
-
-## [0.75.0] - 2026-08-14
-
-### Added
-
-- A real counterparty organization mentioned for the first time now
-  gets auto-created into the corporate hierarchy, not left permanently
-  unresolved. Synthetic regression fixtures prove the first-mention gap.
-  An LLM proposes a
-  Group/Company/Plant placement from context; a real new
-  `corporate_entity` row is only created once the proposal is
-  search-corroborated (reusing the existing Searxng verification
-  client, no new search integration). Auto-created rows get a
-  deterministic `AUTO-`-prefixed code, kept structurally separate from
-  the real login corp-code namespace. Wired into both Keyman
-  affiliation resolution and R&R organization-actor resolution.
-
-## [0.74.0] - 2026-08-14
-
-### Added
-
-- R&R team and organization actors now get a shared identity across
-  posts, not just per-post free text -- the same "설계팀" (design
-  team) named in ten posts resolves to one `cataloged_team` row
-  (identity key: team name + parent org, since a bare team name is not
-  by itself identifying), and an organization actor resolves against
-  the existing `corporate_entity` catalog. Each resolved actor gets a
-  real Knowledge Graph mention edge (`edge_mention_team`,
-  `edge_team_affiliation`, `edge_mention_organization`), so extraction
-  results now genuinely become cross-post lineage clues instead of
-  per-post islands. A person R&R actor is opportunistically joined to
-  an existing Keyman-cataloged person by name (documented gap: R&R does
-  not yet originate new person identities itself -- see ADR 0009).
-
-## [0.73.0] - 2026-08-14
-
-### Fixed
-
-- Image OCR/caption parsing no longer discards a real vision response
-  just because its formatting was close but not exact (bolded labels
-  like `**TEXT:**`, reordered labels, or a missing TAGS line) --
-  observed live against synthetic embedded-image fixtures in the synthetic regression batch
-  in format-variation fixtures. Fields are now recovered independently; only a
-  response with neither TEXT nor CAPTION content is treated as
-  unusable. A strict format mismatch was silently producing the same
-  "[image: content unavailable]" placeholder as a genuinely unavailable
-  vision channel, discarding real, already-paid-for content.
-
-## [0.72.0] - 2026-08-14
-
-### Added
-
-- Abbreviated/slang organization names (e.g. "AGP") are now resolved
-  to their canonical name ("Aurora Grid Power") via LLM context, then
-  cross-verified against external search before being trusted -- new
-  `lineageweave/organization_name_resolution.py`, reusing the existing
-  Searxng verification client rather than a second web-search
-  integration. Cached in a new `organization_name_resolution` table
-  keyed by the raw name.
-- Wired into Keyman affiliation ingestion (both the API path and the
-  offline synthetic-batch script): a search-corroborated resolution feeds
-  `resolve_corporate_entity`, an unverified one leaves the raw name
-  unchanged.
-
-### Fixed
-
-- The offline synthetic-batch script's own re-implementation of Keyman
-  affiliation persistence was missing `role_title` entirely (a stale
-  copy that predated that feature) -- fixed alongside this change.
-
-## [0.71.0] - 2026-08-14
-
-### Added
-
-- VOC evidence now quotes the extractive excerpt under the counterparty
-  it names. After `make seed`, Northridge Grid in VOC evidence shows
-  the sentence that mentions it, instead of a detached excerpt list
-  above the names. Unassigned excerpts stay in the list; a post with
-  no named organization still says so.
-
-## [0.70.0] - 2026-08-14
-
-### Added
-
-- R&R's `actor_type_code` gains `prov_team`, a meso-level actor type for
-  a named sub-unit of a company (e.g. "설계팀"/design team) -- distinct
-  from both a person and the company itself. Grounded in the W3C
-  Organization Ontology's `org:OrganizationalUnit` (Reynolds, 2014).
-- A team actor now requires `affiliated_organization_name` in the same
-  way a person actor does -- a team's own name never answers "which
-  company."
-- R&R badge shows a distinct "Team" label/color, not the prior binary
-  Person/Organization ternary (which would have mislabeled a team as
-  "Organization").
-
-## [0.69.0] - 2026-08-14
-
-### Added
-
-- Keyman extraction now captures a stated job title/position
-  (`PersonMention.job_title`), since two different real people can
-  share a name and a title is real evidence for telling them apart.
-  Persisted to `person_affiliation.role_title` (an existing schema
-  column, previously never populated) and a new
-  `cataloged_person.last_known_job_title` for a title stated without a
-  named organization to attach it to.
-- `_upsert_person` no longer blindly merges a same-name+side match: a
-  genuinely conflicting stated title creates a fresh person row instead
-  of reusing one, verified by a real test with two posts naming the
-  same name and different titles.
-- Keyman panel shows the person's title next to their name. After
-  `make seed`, Ada West is "Account manager" and Priya Nair is
-  "Procurement lead" so the title is visible without a live LLM.
-
-### Fixed
-
-- `scripts/seed_demo_data.py` no longer embeds the local Keycloak admin
-  password. `make seed` still supplies the compose default via
-  `KEYCLOAK_ADMIN_PASSWORD`; a direct script run requires that env var
-  or `--keycloak-admin-password`.
-
-## [0.68.0] - 2026-08-14
-
-### Changed
-
-- R&R's named actor is no longer forced into a person slot. A
-  business post can name an organization acting in its own name
-  ("당사," "Demo Corp"), not an individual --
-  `RoleResponsibility.actor_name` (renamed from `person_name`) now
-  carries `actor_type_code` (Person / Organization, W3C PROV-O
-  grounded) and an LLM-inferred `affiliated_organization_name` for
-  person actors. The popup's R&R list shows a Person/Organization
-  badge and the inferred affiliation; only a person actor still links
-  to the Keyman panel. See ADR 0006.
-
-## [0.67.0] - 2026-08-14
-
-### Added
-
-- Keyman affiliation names that resolve to a cataloged org start
-  the same related-node walk as a related-entity click. After
-  `make seed`, Demo Corp next to Ada West opens
-  `GET /api/corporate-entities/{id}/related`. Unresolved affiliation
-  names stay text -- never a guessed neighborhood.
-
-## [0.66.0] - 2026-08-14
-
-### Added
-
-- Classified counterparty names that resolve to a cataloged org start
-  the same related-node walk as an affiliate-tree org click. After
-  `make seed`, Demo Corp in Counterparties opens
-  `GET /api/corporate-entities/{id}/related`. Unresolved names
-  (Northridge Grid) stay text -- never a guessed neighborhood.
-
-## [0.65.0] - 2026-08-14
-
-### Added
-
-- Resolved affiliate-tree organizations start the same related-node
-  walk as a Keyman org click. After `make seed`, Demo Corp / Test
-  Corp in the tree open `GET /api/corporate-entities/{id}/related`.
-  Unresolved names stay text -- never a guessed neighborhood.
-
-## [0.64.0] - 2026-08-14
-
-### Added
-
-- Related corporate-entity nodes in the Keyman walk are the same
-  RWR as a person. Clicking `Demo Corp` / `Test Corp` loads
-  `GET /api/corporate-entities/{id}/related` so the buyer can
-  continue the walk from the org. An entity with no visible
-  affiliation stays 403 -- never a guessed neighborhood.
-
-## [0.63.0] - 2026-08-14
-
-### Added
-
-- A third seeded Ask question, "What is the next commitment?", on
-  every fixture that already answers the first two chips. A-100
-  names Send Northridge Grid the revised quote due 2026-01-12;
-  B-200 names the Westfield specification due 2026-01-14; Riverbend
-  names its calendar ticket due 2026-01-09; rec-006 says it has no
-  open commitment. After `make seed` the popup chips connect Ask
-  to the same dated tickets home Calendar already lists.
-
-## [0.62.0] - 2026-08-14
-
-### Added
-
-- VOC evidence counterparties that already appear on the affiliate
-  tree are the same Keyman walk. Clicking `Northridge Grid` loads
-  related nodes for Priya Nair so the buyer does not have to find
-  the same org again under Affiliate tree.
-
-### Security
-
-- Verification badges only become links for `http:` / `https:`
-  evidence URLs. A `javascript:` or `data:` value stays a non-link
-  badge.
-
-## [0.61.0] - 2026-08-14
-
-### Changed
-
-- Activity event-type badges show Ticket created / Status changed /
-  Commitment derived instead of snake_case codes. Unknown types stay
-  as the raw code; the badge never invents a name.
-
-## [0.60.0] - 2026-08-14
-
-### Changed
-
-- Activity-feed `ticket_status_changed` summaries use the lookup
-  label (`In progress`, `Closed`) instead of the raw code. Same
-  `common_lookup_value` labels the popup select, Calendar, and
-  period-report members already show.
-
-## [0.59.0] - 2026-08-14
-
-### Added
-
-- Period-report members show the ticket status lookup label next to
-  the title and due date. After `make seed` the A-100 pricing
-  follow-up reads Open plus `due 2026-01-12`, same label the
-  Calendar and popup select already use.
-
-## [0.58.0] - 2026-08-14
-
-### Changed
-
-- After an unmatched Verify 503, the Verify against web search
-  button is hidden the same way as Evaluate / Extract / Derive.
-  Seeded counterparty badges stay; the button that can only 503
-  again does not. Still never fabricates a search result.
-
-## [0.57.0] - 2026-08-14
-
-### Added
-
-- Home Calendar rows show the ticket status lookup label next to the
-  due date. After `make seed` the A-100 pricing ticket reads Open
-  plus `due 2026-01-12`, same label the popup select already uses.
-
-## [0.56.0] - 2026-08-14
-
-### Added
-
-- A second seeded Ask question, "Who is involved?", on every fixture
-  that already answers "What happened between these events?" A-100
-  names Ada West and Priya Nair; B-200 names Jordan Hale; rec-006
-  and Riverbend say they do not name a Keyman. After `make seed`
-  the popup shows two chips instead of one canned prompt.
-
-## [0.55.0] - 2026-08-14
-
-### Added
-
-- Ticket status on the popup select shows `common_lookup_value`
-  labels (`Open`, `In progress`, `Closed`) instead of raw codes.
-  Codes stay on the payload. A missing lookup falls back to the code.
-
-## [0.54.0] - 2026-08-14
-
-### Changed
-
-- After an unmatched 503, Evaluate, Extract Keymen, and Derive
-  commitment are hidden the same way as free-text Ask. Seeded
-  evaluation, Keyman, and ticket rows stay; the buttons that can
-  only 503 again do not. Still never fabricates a live LLM result.
-
-## [0.53.0] - 2026-08-14
-
-### Added
-
-- `make seed` opens a ticket on the B-200 specification-revision
-  fixture. A B-200 report-member click now shows "Send Westfield
-  Power the revised specification" (due 2026-01-14) instead of
-  "No tickets yet," and home Calendar lists it next to the A-100
-  pricing ticket.
-
-## [0.52.0] - 2026-08-14
-
-### Changed
-
-- After an unmatched Ask 503, the free-text input is hidden. Seeded
-  answers stay on screen with "Only seeded questions can be answered
-  without an orchestrator." Still never fabricates a live reply.
-
-## [0.51.0] - 2026-08-14
-
-### Added
-
-- Home Calendar lists the seeded A-100 pricing ticket due 2026-01-12.
-  After `make seed`, GET /api/calendar includes "Send Northridge Grid
-  the revised quote" next to the Riverbend commitment, so the buyer
-  sees the same dated ticket already shown on the period-report row.
-
-## [0.50.0] - 2026-08-14
-
-### Added
-
-- Seeded fixture summaries include R&R from the Event Lineage cast.
-  An A-100 DAG click now shows Ada West / Priya Nair, and a B-200
-  click shows Jordan Hale -- not an empty R&R list. rec-006 and the
-  calendar commitment stay empty because they have no cast people.
-  An R&R name that matches a Keyman opens the same related-node walk
-  as the Keyman and affiliate-tree buttons.
-
-## [0.49.0] - 2026-08-14
-
-### Added
-
-- Period-report members show the earliest open ticket title next to
-  the due date. The home-page report list now reads "Send Northridge
-  Grid the revised quote" plus `due 2026-01-12` on the A-100 pricing
-  follow-up, so the buyer sees the seeded ticket before opening the
-  post.
-
-## [0.48.0] - 2026-08-14
-
-### Added
-
-- Period-report members include the earliest open ticket due date.
-  The home-page report list shows `due 2026-01-12` on the A-100
-  pricing follow-up so the buyer sees the seeded ticket before
-  opening the post.
-
-## [0.47.0] - 2026-08-14
-
-### Added
-
-- Seeded Ask answers for the demo public post, A-100/B-200 Event
-  Lineage fixtures, and the Riverbend calendar post. `GET`/`POST
-  /api/posts/{id}/chat` read `post_chat_result` first so the popup
-  transcript answers "What happened between these events?" without an
-  orchestrator. An unmatched question still 503s -- never a fabricated
-  live reply.
-
-## [0.46.0] - 2026-08-14
-
-### Changed
-
-- Evaluate, Extract Keymen, Derive commitment, and Verify show the
-  same kind of empty state as chat when the backing service is 503
-  (`Evaluation unavailable…`, `Keyman extraction unavailable…`,
-  `Commitment derivation unavailable…`, `Verification unavailable…`)
-  instead of `HTTP 503`. Still never fabricates a result.
-
-## [0.45.0] - 2026-08-14
-
-### Changed
-
-- Chat 503 (orchestrator unset) shows
-  `Chat unavailable (LLM orchestrator not configured).` instead of
-  `HTTP 503`. Still never fabricates an answer.
-
-## [0.44.0] - 2026-08-14
-
-### Added
-
-- `make seed` XADDs a `ticket_created` event onto Valkey for the
-  seeded A-100 and calendar tickets. A report-member click now shows
-  "Ticket created: Send Northridge Grid the revised quote" instead of
-  "No activity yet."
-
-## [0.43.0] - 2026-08-14
-
-### Added
-
-- VOC evidence counterparties show the same verification badge as the
-  Counterparties panel (`Not yet checked` / `Corroborated` plus the
-  evidence URL when present). The extractive quote is not an unchecked
-  second claim.
-
-## [0.42.0] - 2026-08-14
-
-### Added
-
-- `make seed` opens tickets on the A-100 pricing follow-up and
-  delivery-schedule fixtures. A report-member click now shows
-  "Send Northridge Grid the revised quote" (due 2026-01-12) instead
-  of "No tickets yet."
-
-## [0.41.0] - 2026-08-14
-
-### Changed
-
-- Period-report members with Event Lineage, Keyman, and evaluation
-  sort ahead of dummy high/low band rows. A-100 fixtures score on the
-  high process unit and B-200 on the low unit, so a member click opens
-  a reconstruct post that already has those panels.
-
-## [0.40.0] - 2026-08-14
-
-### Added
-
-- Seeded period reports fold A-100/B-200 Event Lineage fixtures and
-  the Riverbend calendar post (already carrying IRT cells) into the
-  shared-metric 2026-W02 bank. The comparison strip can open those DAG
-  posts; dummy high/low band posts remain for the PU ranking test.
-
-## [0.39.0] - 2026-08-14
-
-### Added
-
-- Affiliate-tree people are the same Keyman walk as the Keyman panel.
-  Clicking `Ada West` or `Priya Nair` on the tree loads related nodes
-  (`GET /api/keymen/{id}/related`) so the buyer does not have to find
-  the same name again under Keyman.
-
-## [0.38.0] - 2026-08-14
-
-### Added
-
-- Seeded Keyman and VOC counterparty for B-200 Event Lineage fixtures
-  (Jordan Hale / Westfield Power, Voice of Market). Clicking a B-200
-  DAG node is no longer an empty Keyman or VOC panel. A-100 and
-  rec-006 casts are unchanged.
-
-## [0.37.0] - 2026-08-14
-
-### Added
-
-- Seeded Keymen, affiliate-tree leaves, and VOC counterparties for
-  every A-100 proj-alpha Event Lineage fixture and the calendar
-  Riverbend post. Clicking a DAG node now shows Ada West / Priya Nair,
-  Northridge Grid, and an extractive excerpt without a live
-  orchestrator. rec-006 stays uncast so it remains its own root. A
-  post with no named org still has empty VOC excerpts -- never a
-  fabricated quote.
-
-## [0.36.0] - 2026-08-14
-
-### Added
-
-- `make seed` writes constructed IRT rubric cells for the demo public
-  post, A-100/B-200 Event Lineage fixtures, and the calendar
-  commitment. The Post quality panel shows those categories after a
-  fresh stack -- not "Not yet evaluated." Cells are title-derived, not
-  an LLM judge; thetas still come only from `calibrate_period_report`.
-
-## [0.35.0] - 2026-08-13
-
-### Added
-
-- Post list and popup show `common_lookup_value` labels for VOC type
-  and visibility (`Voice of Customer`, `Public`) instead of raw codes.
-  Codes stay on the payload. A missing lookup falls back to the code.
-
-## [0.34.0] - 2026-08-13
-
-### Added
-
-- Seeded Korean summaries for every A-100/B-200 reconstruct fixture and
-  the calendar commitment post. Clicking an Event Lineage DAG node or a
-  calendar row now serves `GET /api/posts/{id}/summary` from
-  `post_summary_result` without a live orchestrator. A non-fixture post
-  with no stored row still 503s -- never a fabricated summary.
-
-## [0.33.0] - 2026-08-13
-
-### Added
-
-- Home-page grouping comparison strip: `GET /api/reports/compare/{period}`
-  returns every process unit, corporate entity, and thread group scored
-  on the shared FIPC metric. Rebuild now writes all three kinds.
-  Clicking a row switches the Period reports grouping. Seed assigns
-  high posts to thread A-100 and low posts to B-200 so the strip is
-  not empty after `make seed`.
-
-## [0.32.0] - 2026-08-13
-
-### Added
-
-- CAT item-information selection on the shared FIPC bank: after a
-  group is scored, `information_polytomous` ranks rubric items by
-  Fisher information at that group's mean θ (Lord, 1980). Rank 1 is
-  the max-info pick -- a high PU and a low PU can be measured by
-  different items on the same bank. Rankings persist to
-  `report_item_information` and the Period reports panel labels
-  `CAT: sales-lead I=0.70`. Information comes from fast-mlsirm's
-  Rust GRM/GPCM curves, never a hand-rolled I(θ).
-
-## [0.31.0] - 2026-08-13
-
-### Added
-
-- Shared-metric FIPC: the first period free-calibrates one item bank
-  on the pooled posts; every process unit (and other grouping) is
-  then scored on that bank so a high PU and a low PU stay comparable.
-  Independent per-group refits would each re-center near 0. The
-  Period reports panel labels first-period groups `shared metric`.
-
-## [0.30.0] - 2026-08-13
-
-### Added
-
-- Persisted Korean summaries (`post_summary_result` /
-  `post_summary_event` / `post_summary_role`). `GET /api/posts/{id}/summary`
-  returns a stored row first so a seeded demo popup is not empty when
-  the LLM orchestrator is off. Live derivation still writes through
-  the same tables. Seed writes a synthetic summary for the demo public
-  post. A missing stored row and a missing LLM stay 503 -- never a
-  fabricated summary.
-
-## [0.29.0] - 2026-08-13
-
-### Added
-
-- True cross-period FIPC linking: the first week for a grouping
-  free-calibrates and persists `report_item_parameter`; later weeks
-  EAP-score on those fixed item parameters so mean θ is comparable
-  across weeks (Kim, 2006). An independent all-high week would
-  re-center near 0; the linked week stays high on the reference
-  metric. `GET /api/reports/{grouping}` lists the trend; the Period
-  reports panel shows `vs 2026-W02: +0.92`. Seed writes a constructed
-  2026-W02 reference and 2026-W03 all-high follow-up.
-
-## [0.28.2] - 2026-08-13
-
-### Added
-
-- Period reports lists each member post with its fitted EAP θ and
-  opens that post on click -- the same pattern as the calendar.
-
-## [0.28.1] - 2026-08-13
-
-### Added
-
-- `make seed` inserts constructed high/low IRT rows for 2026-W02 and
-  persists a real FIPC report, so the Period reports panel is not empty
-  after a fresh stack. Thetas still come only from
-  `calibrate_period_report`.
-
-## [0.28.0] - 2026-08-13
-
-### Added
-
-- ADR 0003 slice 3: `lineageweave/period_report.py` fits GRM and GPCM
-  with `fast_mlsirm.fit_polytomous`, EAP-scores posts, and picks the
-  model via `fixed_item_calibration_diagnostics`. Scores persist to
-  `report_period_score` / `report_member_score`.
-  `GET /api/reports/{grouping}/{period}` and `POST .../rebuild`
-  (`post_admin`) plus a home-page Period reports panel. The accuracy
-  test constructs high vs low category groups and asserts the fitted
-  mean θ ranks them correctly -- no hardcoded score.
-
-## [0.27.0] - 2026-08-13
-
-### Added
-
-- ADR 0003 slice 2: a post can be LLM-as-a-Judge evaluated through
-  `fast-mlsirm`'s `ContextualOrchestratorJudge`. Scores persist only
-  via `LLMJudgeResult.to_irt_row()` into `post_evaluation_response`
-  (one row per post per rubric criterion). Null client is unavailable,
-  never a fabricated score.
-- Versioned rubric: constructive stance, negative stance, sales-lead
-  specificity. `POST /api/posts/{id}/evaluate` +
-  `GET /api/posts/{id}/evaluation`, and an "Evaluate post" panel on
-  the post popup.
-
-## [0.26.0] - 2026-08-13
-
-### Added
-
-- External search verification for Knowledge Graph relation inferences:
-  `lineageweave/relation_verification.py`'s `SearxngRelationVerificationClient`
-  checks an LLM-classified counterparty organization/relationship
-  against a self-hosted Searxng instance, catching a hallucinated
-  organization name with zero real-world footprint (FEVER grounding:
-  Thorne, Vlachos, Christodoulopoulos, & Mittal, 2018). A search hit
-  only corroborates when a distinctive org-name token appears in the
-  result host or snippet and the host is not itself a search page --
-  engines echo any query in the title, so "any result" is not evidence.
-- New Docker Compose service `searxng` (`docker/searxng/`), wired into
-  the backend via `SEARXNG_BASE_URL`.
-- `post_counterparty_entity` gained `verification_status_code`
-  (`verify_pending` / `verify_corroborated` / `verify_uncorroborated`),
-  `verification_evidence_url`, `verification_checked_at`
-  (`migrations/0004_relation_verification.sql`). A re-classification
-  resets these back to pending.
-- New `POST /api/posts/{id}/verify-relations` endpoint and a
-  `CounterpartyPanel` UI section (status badge, linked to the evidence
-  URL when corroborated, with a "Verify against web search" action).
-- ADR 0005 documents the design decision.
-
-## [0.25.0] - 2026-08-13
-
-### Added
-
-- Post bodies are normalized before ever reaching an LLM or embedding
-  call: HTML tags are stripped to clean text, per-block formatting
-  (`style`, heading level) is kept as separate `FormattingHint` metadata
-  instead of being embedded, and base64-embedded images are described
-  by a vision-capable model and placed as `[image: caption | text: ocr]`
-  at their original document position instead of being sent raw or
-  dropped (`lineageweave/post_content_normalization.py`, VIPS grounding:
-  Cai, Yu, Wen, & Ma, 2003). Plain-text comparison operators (`qty < 50`)
-  are not treated as HTML.
-- `chunking.py`'s DOM chunker now captures each block's `style`
-  attribute and splits on `h1`-`h6` heading boundaries.
-- New `VISION_MODEL` setting; `extract-keymen`, post summary, commitment
-  derivation, and chat source retrieval (including linked posts pulled
-  in as RAG context) all normalize `post_body` before the LLM sees it.
-  The raw post-detail read (`GET /api/posts/{id}`) is intentionally
-  left untouched so the frontend still renders the post as-authored.
-
-## [0.24.0] - 2026-08-13
-
-### Added
-
-- Related Keyman nodes are now walkable: a related post opens that
-  post, a related person loads *their* RWR neighbourhood. Dead text
-  next to "Related to Ada West" is no longer a dead end.
-- Chat citation chips show the source post's title
-  (`cited_posts` from `POST /api/posts/{id}/chat`) instead of a
-  truncated UUID. `cited_post_summaries` drops unknown ids rather than
-  inventing a label.
-
-## [0.23.0] - 2026-08-13
-
-### Added
-
-- Keyman side and corporate-entity level now carry lookup labels from
-  `common_lookup_value` (`person_side_label`, `entity_level_label`).
-  The popup shows "Ada West (Our side)" and "Demo Corp (Company)"
-  instead of raw `our_side` / `company` / `plant` codes. Missing lookup
-  rows still fall back to the code, never an invented name.
-
-## [0.22.0] - 2026-08-13
-
-### Added
-
-- Related Keyman nodes now carry the ontology class (`ontology_iri`,
-  `ontology_label`) from `docs/ontology/lineageweave-kg.ttl`, so the
-  popup shows "Priya Nair (Person)" instead of the raw `node_person`
-  lookup code. Missing terms stay unlabeled rather than inventing a
-  string -- same missing-vs-negative rule as every Null channel.
-  `GET /api/keymen/{id}/related` is the wire; the frontend falls back
-  to `node_type_code` only when the ontology has no term.
-
-### Fixed
-
-- `frontend/package.json` was invalid JSON after #24 dropped the
-  comma on the version line, so `pnpm run test` / CI frontend could
-  not even parse the manifest.
-
-## [0.21.2] - 2026-08-13
-
-### Added
-
-- The Knowledge Graph is now grounded in a real, machine-checked
-  Ontology and Semantic Layer (see
-  [ADR 0004](docs/adr/0004-knowledge-graph-ontology.md)) -- the
-  brief's latest revision required this everywhere the KG is used
-  (Keyman traversal, corporate hierarchy tree, entity-relationship
-  classification, indirect lineage linking, in-popup chat evidence).
-  `docs/ontology/lineageweave-kg.ttl`: a real OWL 2 / RDFS / SKOS
-  ontology (Turtle) formalizing `knowledge_graph_edge`'s node/edge
-  vocabulary and the `entity_relationship_type`/`person_side`/
-  `corporate_entity_level` controlled vocabularies -- classes with
-  declared subclass relations, object properties with
-  `rdfs:domain`/`rdfs:range`, and the corporate-level hierarchy
-  (Group -> Company -> Plant) as a proper SKOS concept scheme.
-  `lineageweave/ontology.py` parses it with `rdflib` and exposes the
-  vocabulary as importable IRI constants.
-- `tests/test_ontology.py`: a real round-trip check, not just "does
-  the file parse" -- every lookup code `scripts/seed_demo_data.py`
-  actually seeds (for the categories this ontology covers) must have
-  a matching ontology term, and vice versa, so the ontology cannot
-  silently drift from the relational schema's real vocabulary.
-
-## [0.21.1] - 2026-08-13
-
-### Fixed
-
-- `test_seed_calendar_commitment_surfaces_on_get_calendar` drives the
-  same `_seed_demo_calendar_commitment` helper `make seed` uses, then
-  `GET /api/calendar`, and asserts the Riverbend ticket is due
-  2026-01-09. Without this, a seed helper that wrote the wrong date
-  would only fail in a live compose stack.
-
-## [0.21.0] - 2026-08-13
-
-### Added
-
-- `make seed` now inserts the synthetic Riverbend commitment post
-  (`fixtures.ambiguous_commitment_post`, created_at 2026-01-05) plus
-  one open dated ticket (due 2026-01-09). Fresh stacks show a row on
-  the home-page Calendar instead of an empty panel that made 0.18.0's
-  surface look unfinished. Re-seed is idempotent. The empty-state copy
-  tells the buyer how to populate the calendar (Derive, or a ticket
-  with a due date) when there really are none.
-
-## [0.20.0] - 2026-08-13
-
-### Added
-
-- The post-detail Event Lineage panel now draws the same git-branch SVG
-  as the home page, scoped to that post's reconstruct group
-  (`subgraphForPost`). Opening an A-100 post shows the `rec-002` fork
-  and keeps B-200 out; the direct/indirect lists stay as the
-  why-linked distinction. Tests drive `subgraphForPost` on the
-  designed A-100 fixture.
-
-## [0.19.0] - 2026-08-13
-
-### Added
-
-- `fast-mlsirm` as a pinned git dependency (`pyproject.toml`) -- the
-  first of three staged slices toward weekly/monthly PU/team/project
-  reports (see [ADR 0003](docs/adr/0003-fast-mlsirm-report-integration.md)).
-  It already implements the brief's LLM-as-a-Judge -> IRT-row ->
-  Fixed-Item Parameter Calibration pipeline, provider-neutrally against
-  contextual-orchestrator; reused rather than reimplemented. No
-  product behavior change in this slice -- infra only.
-- `requires-python` moved from `>=3.10` to `>=3.12` to match
-  `fast-mlsirm`'s own floor. `backend/Dockerfile`'s build stage gained
-  a pinned, non-interactive `rustup` install (`build-essential` for
-  the linker), since `fast-mlsirm` ships a PyO3/maturin Rust core with
-  no fallback wheel. The pytest CI job installs the same toolchain
-  before `pip install -e ".[backend]"`. Verified the compiled
-  extension actually loads (not the NumPy parity fallback) both
-  locally and in a freshly built `backend` Docker image. TEPP stays
-  the temporal/event measurement platform -- this pin is the IRT
-  compute library, not a fork of TEPP.
-
-## [0.18.0] - 2026-08-13
-
-### Added
-
-- Customer commitment derivation and a calendar/to-do surface -- the
-  brief's "이슈 관리는 To Do 와 캘린더로 자동 등록" and "고객과의
-  약속에 관해서 LLM 자동 도출" items, unified into one design: a
-  derived commitment *is* the ticket that shows up on the calendar,
-  not a separate concept. `lineageweave/commitment_extraction.py`:
-  pluggable LLM client that decides whether a post contains a genuine
-  customer commitment (a promise with a deadline, not just any event)
-  and resolves relative phrases ("by next Friday") against a supplied
-  reference date. `POST /api/posts/{post_id}/derive-commitment`
-  (`post_admin`) runs it and, when found, persists an `issue_ticket`
-  with `due_date`/`commitment_summary` set; `has_commitment: false` is
-  a normal 200, not an error, since most posts have no commitment.
-  `GET /api/calendar` lists every dated ticket the account may see,
-  soonest first, ABAC-filtered per row like every other cross-post
-  endpoint.
-- `issue_ticket` gained `due_date` and `commitment_summary` columns
-  (folded into `migrations/0001_initial_schema.sql`, plus
-  `migrations/0003_ticket_commitment_calendar.sql` for upgrading an
-  existing volume, matching the `0002` precedent).
-- Frontend: a "Derive commitment" action in the ticket panel, due
-  dates shown on ticket rows, and a Calendar panel on the product home
-  page listing upcoming commitments across every visible post,
-  clickable through to the source post.
-- Verified with a real LLM call through a locally-launched
-  contextual-orchestrator instance: `ambiguous_commitment_post`'s
-  relative "by next Friday" deadline resolved to the correct absolute
-  date against a supplied reference date, and a second sentence that
-  merely *looks* date-adjacent (a past event) was correctly not
-  treated as a commitment.
-
-### Fixed
-
-- `due_date` was a `timestamptz`. Binding a Python `date` becomes
-  midnight in the session timezone, then `.date()` on the UTC
-  (or local) datetime shifts the calendar day -- "2026-01-09"
-  came back as "2026-01-08". The column is now a `date`, matching
-  the LLM's YYYY-MM-DD. A malformed string still surfaces as 422.
-- Derive used wall-clock `now()` as the reference date. Relative
-  phrases in a historical post ("by next Friday") must resolve against
-  the post's `created_at` (TimeML document creation time), or the
-  calendar entry lands on the Friday after the click, not the Friday
-  the commitment was made.
-- Re-deriving the same post stacked a second open ticket. The persist
-  path now upserts the existing open commitment ticket.
-- `GET /api/calendar` included closed tickets, so finished work still
-  looked upcoming. Closed rows are excluded; a dated ticket created
-  through the regular ticket API still appears while it is open.
-- The "Derive commitment" button was shown to accounts without
-  `post_admin` and then 403'd. It is now gated the same way as Extract
-  Keymen. Calendar rows also show the source post title.
-- Ticket SQL stopped interpolating a column-list constant via f-string
-  (`asyncpg-sqli` Semgrep). The column list is a static literal in
-  each query; values stay parameterized.
-
-## [0.17.0] - 2026-08-13
-
-### Added
-
-- Valkey as a real event queue -- not a cache, not a second database. It
-  had been running in `docker-compose.yml` since Phase 1 with nothing in
-  the codebase ever publishing or reading from it (the brief's "Event
-  Queue, not MQ" requirement was unfulfilled infrastructure).
-  `backend/app/activity_stream.py`: ticket create/status-change
-  mutations `XADD` an event onto a per-post stream
-  (`activity:{post_id}`, approximately trimmed to the most recent 1000
-  entries); `GET /api/posts/{post_id}/activity` reads it straight back
-  with `XREVRANGE` -- no consumer group, no background worker, the
-  smallest slice where Valkey is actually load-bearing.
-- Frontend: an Activity panel in the popup (list + manual Refresh),
-  genuinely wired to the new endpoint.
-- Verified through the real Docker Compose network end to end (not just
-  `pytest`): created and patched a ticket via the actual backend
-  container, confirmed the events on `GET .../activity`, and confirmed
-  directly with `valkey-cli XLEN` against the actual `valkey` container
-  that the stream is real.
-
-## [0.16.0] - 2026-08-13
-
-### Added
-
-- Affiliate tree and VOC evidence on the post-detail popup. The schema
-  already stored a self-referencing `corporate_entity` hierarchy and a
-  `voc_type_code`, but the buyer only saw a raw code and a comma-separated
-  affiliation list.
-  `lineageweave/affiliate_tree.py` builds the ancestor forest of the
-  organizations a post's Keymen actually touch (Bhattacharya & Getoor,
-  2007's candidate-generation stage: an unresolved name is its own root,
-  never a guessed parent). `lineageweave/voc_evidence.py` quotes the
-  sentence that names a classified organization (ACE mention extent;
-  Doddington et al., 2004) and returns nothing when the name is absent.
-- Backend: `GET /api/posts/{post_id}/affiliate-tree` and
-  `GET /api/posts/{post_id}/voc-evidence`, same RBAC+ABAC gate as every
-  other post endpoint. Seed now inserts Demo Group → Demo Corp so the
-  tree is visible on a freshly seeded stack.
-- Frontend: Affiliate tree panel, VOC excerpt block, clickable Keyman
-  (loads `GET /api/keymen/{id}/related`), Event Lineage links that open
-  the linked post, and Extract Keymen for `post_admin`.
-
-## [0.15.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 5: issue ticket management -- the one explicit
-  product-brief item with a schema table (`issue_ticket`) but no
-  implementation until now. `backend/app/issue_ticket_ingestion.py`:
-  CRUD helpers, deliberately not a pluggable-LLM channel -- ticket
-  status is a closed enum in `common_lookup_value`.
-  `GET`/`POST /api/posts/{post_id}/tickets`,
-  `PATCH /api/tickets/{issue_ticket_id}` -- same RBAC (`post_read` for
-  reads, `post_admin` for writes) + ABAC as every other post-scoped
-  endpoint. `PATCH` resolves the ticket's owning post first and
-  ABAC-checks that. An invalid `ticket_status_code` hits the real FK
-  and surfaces as 422.
-- Frontend: `IssueTicketPanel` in the popup -- list, create, and inline
-  status-update.
-
-### Fixed
-
-- `frontend/Dockerfile`: pid-file sed now matches `/run/nginx.pid`
-  (the alpine image's real path), so non-root `USER nginx` can start.
-
-## [0.14.0] - 2026-08-13
-
-### Added
-
-- Product home page renders the reconstruct DAG as a git-branch SVG
-  (same layout language as `web/index.html`), grouped by reconstruct's
-  `thread_group_key`. Branch points are orange, isolated roots stay
-  visible, and a node click opens that post. `GET /api/lineage` now
-  includes `group` from the same `reconstruct_group_key()` rebuild uses,
-  so the UI cannot split A-100 / B-200 differently than reconstruct.
-- `Rebuild lineage` on the home page for `post_admin` accounts
-  (`POST /api/lineage/rebuild`). Viewers still only see the graph.
-
-## [0.13.1] - 2026-08-13
-
-### Fixed
-
-- Rebuild no longer derives reconstruct `group_key`/`secondary_key` from
-  process unit or voc type. Those collapsed A-100/B-200 and
-  proj-alpha/proj-beta and lost the designed `rec-002` fork.
-  `source_post` now persists `thread_group_key` and
-  `secondary_grouping_key`; seed writes fixture `occurred_at` into
-  `created_at`. `test_rebuild_lineage_recovers_the_a100_fork` drives
-  rebuild on the same A-100+B-200 rows seed inserts.
-
-## [0.13.0] - 2026-08-13
-
-### Added
-
-- `POST /api/lineage/rebuild` (`post_admin`): runs `reconstruct()` over
-  every `source_post` and writes `post_lineage_edge`. Grouping keys are
-  persisted on the post (`thread_group_key` / `secondary_grouping_key`).
-- `GET /api/lineage`: ABAC-filtered `{nodes, edges}` in the same shape as
-  the stdlib demo server, so the product UI can show the reconstructed
-  DAG. The React home page now lists those parent→child edges.
-- `backend/tests/test_api.py::test_rebuild_lineage_recovers_the_a100_fork`
-  inserts the designed A-100 fixture as real `source_post` rows, rebuilds,
-  and asserts the fork is visible on both the graph and per-post lineage.
-
-## [0.12.0] - 2026-08-13
-
-### Added
-
-- `lineageweave/lineage_persistence.py`: flattens `reconstruct()` trees
-  into the `(parent, child, fused_score)` rows `post_lineage_edge`
-  stores. `tests/test_lineage_persistence.py` proves the designed A-100
-  fork (`rec-002` → `rec-003` and `rec-004`) is in that contract, and
-  that unrelated `rec-006` is not forced onto a parent.
-- `scripts/seed_demo_data.py` now inserts the synthetic A-100 / B-200
-  fixture posts and persists those reconstruct edges, so
-  `GET /api/posts/{id}/lineage` (the Event Lineage panel) is not empty
-  on a freshly seeded demo stack.
-
-## [0.11.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 4: the post-detail popup's remaining panels.
-  `lineageweave/post_summary.py` -- LLM-derived Korean summary, key
-  events (ACE-style, Doddington et al. 2004), and R&R (semantic role
-  labeling, Gildea & Jurafsky, 2002), via contextual-orchestrator.
-  `lineageweave/post_chat.py` -- in-popup chat as retrieval-augmented
-  generation (Lewis et al., 2020): an explicit retrieve step
-  (`backend/app/post_chat_ingestion.py` assembles a post's own content
-  plus its Event-Lineage-linked posts, direct and Knowledge-Graph-
-  indirect, as numbered sources) then a reason-and-cite step (the model
-  answers using only those sources and reports which it drew from).
-- Backend: `GET /api/posts/{post_id}/lineage` (direct vs. indirect links,
-  kept as two separate lists), `GET /api/posts/{post_id}/summary`,
-  `POST /api/posts/{post_id}/chat` -- all RBAC+ABAC-gated the same way as
-  every other post endpoint, verified end to end against a live
-  Postgres + Keycloak + contextual-orchestrator stack (including through
-  the actual Docker-built images, not just the FastAPI TestClient).
-  Caught and fixed a real bug while building the chat's retrieve step:
-  `backend/app/knowledge_graph.py::load_visible_subgraph` only loads
-  edges among an *already-known* post set -- it does not itself discover
-  sibling posts sharing a mentioned person (its only prior caller,
-  `related_for_person`, pre-resolves that full set itself before calling
-  it). `find_linked_post_ids` now does that expansion first; regression-
-  tested (`test_post_chat_cites_a_post_linked_only_via_a_shared_keyman`).
-- Frontend: `frontend/src/App.tsx`'s popup gained real Summary, Event
-  Lineage (direct/indirect visually distinguished), Keyman, Counterparty,
-  and Chat sections, plus a real sliding evidence panel (`EvidencePanel`,
-  CSS animation, not a mock) that opens the cited source post's actual
-  content when a citation chip is clicked. Fixed a real, pre-existing
-  test-infrastructure bug found while adding more tests to the same file:
-  `@testing-library/react`'s automatic per-test cleanup never actually
-  ran (this project's `vite.config.ts` deliberately doesn't set
-  `test.globals`, which auto-cleanup silently depends on), so DOM from
-  one test was bleeding into the next; `src/setupTests.ts` now registers
-  `cleanup` in an explicit `afterEach`.
-- `docs/adr/0002-figma-access-boundary.md`: the referenced Figma frame is
-  genuinely reachable in this environment, but its own cover page is the
-  source organization's real confidential content (and it does not yet
-  contain a frame for this screen) -- the popup is built from the
-  product brief's text instead, named as an explicit, honest gap rather
-  than either faking a "Figma-matched" claim or stalling on it.
-- New citations: See, Liu, & Manning (2017); Gildea & Jurafsky (2002);
-  Lewis et al. (2020).
-
-## [0.10.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 3: `lineageweave/entity_relationship_classification.py`
-  -- LLM classification (via contextual-orchestrator, `mode="route"`) of a
-  named organization's relationship to the post author's org into the
-  product's own six-way vocabulary (`rel_voc`/`rel_vom`/`rel_vop`/
-  `rel_vocc`/`rel_voco`/`rel_vos` -- `rel_`-prefixed because
-  `common_lookup_value.lookup_code` is unique globally across categories
-  and bare `voc`/`vom` were already claimed by `source_post.voc_type_code`'s own
-  category). Proven with a real LLM call against a genuinely hard fixture
-  (`fixtures.ambiguous_entity_relationship_post`): an organization that is
-  both a repeat customer and a newly-competing division in the same post.
-- `lineageweave/corporate_hierarchy_resolution.py` -- similarity-based
-  resolution of a free-text organization name to an existing
-  `corporate_entity` row (Bhattacharya & Getoor, 2007's candidate-
-  generation/blocking stage of collective entity resolution, honestly
-  documented as that stage and not the full joint-inference version).
-  Wired into `backend/app/keyman_ingestion.py`, replacing the exact-
-  case-insensitive-match lookup it had before. `tests/test_corporate_hierarchy_resolution.py`
-  proves it against the same "Acme Group / Acme Electronics Korea / Acme
-  Electronics Gwangju Plant" hierarchy `tests/test_schema.py` already
-  uses: an abbreviation and a trailing legal suffix still resolve to the
-  right entity (not a sibling with a similar name), and a genuinely
-  unrelated organization resolves to `None`, not a guess.
-- `tests/test_indirect_lineage_linking.py`: demonstrates the Knowledge
-  Graph layer finds a real relation `lineageweave.reconstruct` has no
-  mechanism to find at all -- two posts in different `reconstruct.py`
-  groups (structurally never compared against each other) still surface
-  as related once they share a Keyman, via `random_walk_with_restart`.
-- Backend: `GET /api/posts/{post_id}/counterparties`, and
-  `POST /api/posts/{post_id}/extract-keymen` now also classifies and
-  persists each extracted Keyman's affiliated organizations'
-  relationships (`backend/app/entity_relationship_ingestion.py`), all on
-  the same RBAC+ABAC gate as the post/Keyman endpoints. Proven end to end
-  against a live Postgres + Keycloak + contextual-orchestrator stack.
-- `docs/lineage-bi-research-notes.md`: added Zelenko, Aone, & Richardella
-  (2003) (relation extraction); moved Bhattacharya & Getoor (2007) from
-  "staged for later phases" into a real section now that it grounds
-  shipped code.
-
-## [0.9.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 2 continues: `lineageweave/keyman_extraction.py`
-  extracts two-sided Keymen (our-side vs counterparty, 0..N organization
-  affiliations) from a post's title+body. The live client calls
-  contextual-orchestrator with `mode="route"` -- never a raw LLM API.
-  `NullKeymanExtractionClient` stays unavailable rather than inventing
-  mentions.
-- `lineageweave.knowledge_graph.knowledge_graph_edges_for_post` populates
-  the three Phase 2 edge kinds (person<->post mention, person<->corporate
-  entity affiliation, person<->person co-mention) as typed
-  `knowledge_graph_edge` specs. RWR then runs on that graph.
-- Backend endpoints `GET /api/posts/{post_id}/keymen`,
-  `GET /api/keymen/{person_id}/related`, and
-  `POST /api/posts/{post_id}/extract-keymen`, RBAC+ABAC-gated the same
-  way as post detail: a Keyman who is only mentioned on another corp's
-  private post 403s, related-node traversal never returns those hidden
-  posts, and extraction is `post_admin` (a write with a real LLM cost).
-  Persist goes through `backend/app/keyman_ingestion.py` into
-  `cataloged_person` / `person_affiliation` / `post_person_mention` /
-  `knowledge_graph_edge`. Directed RWR sinks teleport remaining walk
-  mass back to the start node so relevance stays a distribution.
-- `fixtures.ambiguous_keyman_post`: a synthetic, non-templated workshop
-  follow-up used by the parser tests and the real-orchestrator Keyman test.
-
-## [0.8.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 2 begins (Phase 1 complete): `lineageweave/knowledge_graph.py`
-  -- random walk with restart (Tong, Faloutsos, & Pan, 2006) giving every
-  node in a Knowledge Graph a continuous relevance score from a starting
-  node, plus `select_related_nodes`'s per-node adaptive relevance-ratio
-  cutoff. No hop-count constant appears anywhere in the algorithm -- the
-  same ratio threshold naturally yields a larger related-set for a
-  well-connected node than a sparse one, which is the actual product
-  requirement ("각 Node와 Node마다 Depth가 다를 수 있다").
-- `tests/test_knowledge_graph.py`: proves this against a synthetic graph
-  built so the correct answer is known by construction -- a disconnected
-  node scores exactly 0, symmetric spokes score identically, and a
-  well-connected "hub" node's adaptive related-set (5 nodes) is
-  measurably larger than a sparse "loner" node's (1 node) under the same
-  threshold.
-- `docs/lineage-bi-research-notes.md`: Tong et al. (2006) moved from
-  "staged for later phases" into a real "Knowledge Graph traversal"
-  section now that it grounds shipped code.
-
-## [0.7.0] - 2026-08-13
-
-### Added
-
-- `frontend/`: React + Vite + TypeScript, pinned Node via `mise.toml`,
-  pnpm via Corepack -- a real client, not mocked and not static HTML.
-  `react-oidc-context` drives an actual Authorization Code redirect
-  through Keycloak; the post list and detail popup call the FastAPI
-  backend over real `fetch()` with the token Keycloak issued.
-  `src/App.test.tsx` covers the login-redirect and fetch-then-render
-  paths (`useAuth` mocked -- the real OIDC round-trip is proven
-  elsewhere, by `scripts/smoke_test_oidc.py` and `backend/tests/test_api.py`).
-- `frontend/Dockerfile` + `nginx.conf`: two-stage build (`pnpm run build`
-  then nginx serving the static bundle) added as docker-compose's fourth
-  service, `VITE_*` config baked in at build time from the same `.env`
-  ports every other service uses. Both stages pin the base image by
-  digest; the runtime stage declares `USER nginx` and listens on 8080
-  so the master process does not need root to bind a port.
-- `backend/app/main.py`: `CORSMiddleware`, scoped to exactly the
-  frontend's origin(s) (`FRONTEND_ORIGINS`), `GET` only, `Authorization`
-  header only -- verified with a real cross-origin preflight + GET against
-  the live stack, not just unit-tested in isolation.
-- `.github/workflows/tests.yml`: added a `frontend` job (lint, test,
-  build) alongside the existing Python `pytest` job.
-
-### Fixed
-
-- Keycloak's `lineageweave-frontend` client (`docker/keycloak/realm-export.json`)
-  now allows both the Vite dev-server origin (`:5173`) and the
-  docker-compose-served frontend's origin (`:15173`) as redirect URIs and
-  web origins -- the login redirect only worked from one of the two
-  before this.
-
-## [0.6.0] - 2026-08-13
-
-### Added
-
-- `backend/`: a FastAPI app connecting directly to PostgreSQL (`asyncpg`,
-  no ORM, no file-backed DB). OIDC bearer-token login verified against a
-  live Keycloak JWKS (`backend/app/auth.py`, fetched via
-  `lineageweave.http_client` rather than PyJWKClient/`urllib`); RBAC
-  (`post_read` permission via role membership) plus row-level ABAC
-  (private `source_post` rows scoped to the requesting account's
-  affiliated corporate entity) enforced on `GET /api/posts` and
-  `GET /api/posts/{post_id}` (`backend/app/main.py`).
-- `backend/tests/test_api.py`: real-integration tests -- a genuine access
-  token from a live Keycloak, verified against a throwaway migrated
-  Postgres database. Proves both the allow and the deny path: a private
-  post scoped to a *different* corporate entity is excluded from the list
-  and 403s on direct fetch; a forged token is rejected; a missing token is
-  401. Skipped unless both a local PostgreSQL and Keycloak are reachable.
-- `scripts/seed_demo_data.py` (`make seed`): seeds synthetic corp/PU/
-  account/`source_post` rows keyed to the *real* subject ids Keycloak's
-  admin REST API reports for the two demo users -- not a locally-fabricated
-  guess at what those ids might be. Talks to Keycloak through
-  `lineageweave.http_client` (`post_form` / `get_json_list`), never
-  `urllib.request.urlopen`.
-- `backend/Dockerfile`: `python:3.12-slim` pinned by digest, runtime
-  `USER appuser` (DS-0002).
-- `migrations/0001_initial_schema.sql`: added `corporate_entity.
-  corporate_entity_code` (unique short code, e.g. `DEMO-CORP-01`) -- the
-  column the login-time `corp_code` claim actually maps to; the original
-  Phase 1 migration (still unmerged) only had the human-readable
-  `entity_name`. Postgres's app database is now auto-migrated with this
-  exact file on first `docker compose up` (`docker/postgres-init/Dockerfile`),
-  so what's tested and what ships never drift apart.
-- docker-compose.yml's default host ports moved off Postgres/Redis/common
-  local-dev ports entirely (15432, 16379, 18080, 18420) -- found during
-  this work that a colliding already-running service on a container's
-  published port can silently answer curl/psql requests instead of the
-  container, with no error; picking non-default ports avoids that
-  ambiguity outright rather than relying on operators noticing.
-
-### Fixed
-
-- JWKS fetch and the demo seeder go through `lineageweave.http_client`
-  instead of `PyJWKClient` / `urllib.request.urlopen`.
-
-## [0.5.0] - 2026-08-13
-
-### Added
-
-- `docker-compose.yml`: PostgreSQL, Valkey, and a real Keycloak OIDC
-  provider, genuinely functional (not a stub/mocked adapter). `make up`
-  brings up all three from a clean checkout; `make smoke` runs
-  `scripts/smoke_test_oidc.py`, which logs in as a synthetic demo user
-  seeded by `docker/keycloak/realm-export.json`, fetches Keycloak's live
-  JWKS, and cryptographically verifies the returned JWT's RS256 signature,
-  issuer, and `corp_code`/`pu_code` custom claims -- a real round-trip
-  proof, not a "the container started" check.
-- `docker/postgres-init/` and `docker/keycloak/`: both services are `build:`
-  targets (Dockerfiles that `COPY` in the keycloak-db init script and the
-  realm seed) rather than bind mounts, so the images are self-contained and
-  reproducible on any Docker host or CI runner.
-- Keycloak stores its own state in a second database (`keycloak`) on the
-  same PostgreSQL instance -- one running database service for the whole
-  stack, no second file-backed store.
-- `.env.example` documents the (already-defaulted) compose variables,
-  including how to remap host ports if 5432/6379/8080 are already taken
-  locally.
-
-### Fixed
-
-- Dockerfiles declare an explicit non-root `USER` (DS-0002) and pin
-  `postgres:16-alpine`, `quay.io/keycloak/keycloak:26.0`, and
-  `valkey/valkey:8-alpine` by digest.
-- `scripts/smoke_test_oidc.py` talks to Keycloak through
-  `lineageweave.http_client` (`get_json` / `post_form`) instead of
-  `urllib.request.urlopen`, so the same `file://` allowlist used by the
-  library clients applies to the OIDC smoke path.
-
-## [0.4.0] - 2026-08-13
-
-### Added
-
-- Milestone 4, Phase 1 begins: LineageWeave's product schema.
-  `migrations/0001_initial_schema.sql` -- a 3NF PostgreSQL schema
-  (snake_case, 2+ word object names) covering accounts (corp/PU code as
-  attributes, not the login key), a shared `common_lookup_value` ENUM
-  table, posts + visibility, ABAC/RBAC, VOC-type and entity-relationship
-  classification, Keyman (`cataloged_person` + N:N `person_affiliation`), a
-  `knowledge_graph_edge` table, `issue_ticket`, and a self-referencing
-  `corporate_hierarchy` via `corporate_entity.parent_entity_id`.
-- `docs/adr/0001-demo-identity-and-data-boundary.md`: the identity/data
-  scope decision for this expansion -- real infrastructure (Postgres,
-  Valkey, a real OIDC provider), synthetic identities and content, because
-  Keyman extraction catalogs real named individuals (including
-  non-consenting external counterparties) and a real production identity
-  provider would re-identify the source organization through account/data
-  structure even with zero literal company-name strings in source files.
-- `tests/test_schema.py`: real-database tests (skipped without a
-  reachable PostgreSQL server) proving the migration applies cleanly, a
-  multi-level corporate-hierarchy recursive query returns the right
-  shape, and an invalid lookup code is genuinely rejected by a foreign
-  key -- caught and fixed a real bug in the process (an accidental
-  `deferrable initially deferred` on one FK silently weakened its
-  integrity check within a transaction).
-- New citations staged for Phase 2/3: Tong et al. (2006, random walk with
-  restart -- Knowledge Graph per-node traversal depth) and Bhattacharya &
-  Getoor (2007, collective entity resolution -- corporate hierarchy).
-
-## [0.3.0] - 2026-08-13
-
-### Fixed
-
-- Embedding, adjudication, and vision clients now POST through a shared
-  `http_client.post_json` helper that allowlists `http`/`https` and never
-  calls `urllib.request.urlopen`. That closes the `file://` read concern
-  Semgrep's `dynamic-urllib-use-detected` rule was flagging on the
-  operator-configured base URLs. HTTPS posts wrap the connected socket
-  with a certifi-backed `SSLContext` instead of constructing
-  `http.client.HTTPSConnection`, so certificate verification is explicit
-  on the Python 3.10+ runtime this project requires.
-
-### Added
-
-- `lineageweave/chunking.py`: semantic-unit chunking so the embedding
-  channel compares meaning-identifiable units instead of whole flattened
-  documents -- `chunk_by_paragraph` (Hearst, 1997, TextTiling subtopic
-  boundaries), `chunk_by_sentence`, `chunk_by_dom` (WHATWG HTML Living
-  Standard sectioning/flow block elements), and `chunk_by_conversation_turn`
-  (RFC 5322 sender/receiver boundaries).
-- `embedding_client.chunked_max_similarity`: chunks two documents, embeds
-  every chunk, and returns the single highest-scoring pair -- the standard
-  passage-retrieval strategy for "a relevant unit is buried in a longer
-  document." Degrades to plain whole-text embedding for any document that
-  chunks to zero or one piece (this project's real short-title dataset
-  behaves exactly as it did before chunking existed).
-- Real-provider test proving chunking works, not just that it type-checks:
-  a short relevant paragraph buried inside a longer synthetic document
-  scored higher via `chunked_max_similarity` than via whole-document
-  embedding, against the live embedding provider.
-- `docs/lineage-bi-research-notes.md`: new "Chunking" section with the
-  four units' grounding and an explicit, honest note that this project's
-  unseen dataset's only free-text field is too short to need chunking in
-  practice -- the module exists for richer content sources (e.g. the raw
-  MHTML artifacts that dataset's records were derived from).
-- `lineageweave/image_content.py`: pluggable vision channel for base64
-  images embedded in DOM content -- real OCR (Li et al., 2023, TrOCR) and
-  object recognition/tagging (Radford et al., 2021, CLIP) via
-  `OpenAiCompatibleVisionClient`, same never-fake-a-missing-channel
-  discipline as the embedding/adjudication clients. `chunk_by_dom` now
-  extracts embedded images as `"image"` chunks interleaved with text
-  chunks in true document order, so an image's position relative to its
-  surrounding text is preserved and reconstructable.
-- Real-provider test proving OCR works, not just that it type-checks: a
-  real PNG generated with real rendered text (not a fixture file) was
-  read back correctly by `OpenAiCompatibleVisionClient` against the live
-  vision-capable model.
-- `docs/image-content-schema.md`: proposed DB schema (snake_case, 2+ word
-  object names) for persisting and searching extracted image content,
-  designed so a text/tag search hit stays traceable to which document and
-  which position produced it, and so the same image (by content hash) is
-  never *stored* twice (the primary key guarantees that part). Avoiding a
-  duplicate vision-provider *call* for two concurrent ingests of the same
-  new image is a separate concern the schema documents but does not solve
-  by itself -- a real write path still needs an atomic claim/lease step.
-
-## [0.2.0] - 2026-08-13
-
-### Added
-
-- ADR-0016 grounding: `docs/lineage-bi-research-notes.md` now cites
-  Doddington et al. (2004, ACE) and Anagnostopoulos et al. (2013, CHRONOS)
-  alongside Allan (2002, TDT), and explains how `reconstruct.py` maps onto
-  TEPP's three-layer event-intelligence separation (mention/instance,
-  calibrated detection, temporal-consistency).
-- `tests/test_real_provider_integration.py`: opt-in (env-var-gated, skipped
-  by default so a credential-free clone stays green) tests proving
-  `OpenAiCompatibleEmbeddingClient` and `ContextualOrchestratorAdjudicationClient`
-  work end-to-end against real providers, not just that their interfaces
-  are satisfiable by a stub.
-
-### Fixed
-
-- `embedding_client.py` / `adjudication_client.py`: both real-provider HTTP
-  clients now build their SSL context from `certifi`'s CA bundle instead of
-  the platform default. Some interpreter distributions (observed: a
-  standalone uv-managed CPython on macOS) don't reliably inherit the OS
-  trust store, which made every real-provider call fail closed with
-  `CERTIFICATE_VERIFY_FAILED` even against a validly, publicly-trusted
-  certificate. Full chain validation still applies -- nothing is weakened,
-  the bundle source just changed.
-
-## [0.1.0] - 2026-08-13
-
-### Added
-
-- Initial prototype: `reconstruct()` pipeline (group → bounded candidate
-  window → multi-channel fusion via RankWeave → tree assembly via
-  ThreadWeave).
-- Four scoring channels: `temporal`, `secondary_key`, `text`
-  (dependency-free stand-in for embedding-cosine similarity), and an
-  optional `llm` channel via a pluggable `AdjudicationClient`.
-- `ContextualOrchestratorAdjudicationClient`, calling
-  [contextual-orchestrator](https://github.com/ContextualWisdomLab/contextual-orchestrator)'s
-  `mode="verify"` with `reasoning_effort="high"`.
-- `OpenAiCompatibleEmbeddingClient` for a real embedding-cosine text
-  channel against any OpenAI-compatible `/v1/embeddings` endpoint.
-- `TeppClient` / `AnalysisRunRequest`: validated wire shape matching
-  [TEPP](https://github.com/ContextualWisdomLab/TEPP)'s published
-  `analysis_run_request_v1.json` schema, pluggable transport (fails closed
-  with `TeppNotAvailable` until TEPP ships a live HTTP endpoint).
-- Minimum fused-score floor so weakly-related records surface as their own
-  root instead of being force-attached to the best of a bad set of
-  candidates.
-- Stdlib HTTP demo server (`lineageweave/server.py`) and a self-contained
-  SVG DAG viewer (`web/index.html`), no build step or external script
-  dependency.
-- Synthetic-only demo dataset (`lineageweave/fixtures.py`).
-- `docs/lineage-bi-research-notes.md`: the literature this design is
-  grounded in, APA 7th.
+  verification) rather than rushed. 553 other tests unaffected (some
+  tests skipped by design in CI).

From 4e4986dd92435ccb6f590242589c73724c3cffd0 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:05:12 -0700
Subject: [PATCH 08/11] fix(changelog): restore historical entries

---
 CHANGELOG.md | 2042 +++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 2030 insertions(+), 12 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index ae29adc82..c8ed1a099 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -6,18 +6,8 @@ All notable changes to this project are documented here. Format follows
 
 ## [Unreleased]
 
-### Added
-
-- Ask Agent citations can now open a focused evidence layer that preserves the
-  current answer while exposing persisted text and image evidence, with
-  Storybook coverage for populated, empty, missing-OCR, null-caption, and
-  blank-caption states.
-
 ### Fixed
 
-- The Ask citation evidence layer now contains forward and reverse keyboard
-  focus while marked modal, gives its modal evidence lists citation-specific
-  accessible names, and falls back to `Untitled image` for blank captions.
 - `make smoke` and `make seed` now run through the locked project `uv`
   environment, so local OIDC and synthetic-data workflows resolve the same
   pinned dependencies as CI.
@@ -81,5 +71,2033 @@ All notable changes to this project are documented here. Format follows
   constraint). Root cause not yet conclusively identified; deferred
   as out of scope for this migration-catchup change (a different
   feature area -- analysis-run/TEPP lifecycle, not R&R/summary/
-  verification) rather than rushed. 553 other tests unaffected (some
-  tests skipped by design in CI).
+  verification) rather than rushed. 553 other tests unaffected.
+
+- `get_or_create_corporate_entity`'s post-lock duplicate-create re-check
+  fuzzy-matched against every cataloged entity, not just an exact
+  concurrent duplicate of the entity being created. A newly-created
+  parent whose name is a prefix of the child now being created (e.g.
+  "Acme" as parent of "Acme Gwangju Plant") scored ~0.7 similarity
+  against that child under the shared 0.6 threshold, so the child was
+  silently bound to its own parent's id instead of getting its own
+  catalog row -- undermining exactly the "통합 고객사 계열 tree AI"
+  (integrated customer affiliate tree) hierarchy the feature exists
+  for. The re-check now requires an exact post-normalization match
+  (`min_similarity=1.0`); real mention resolution against the full
+  candidate set is unchanged. Caught locally by
+  `test_first_mention_of_a_new_counterparty_creates_a_real_corporate_entity`,
+  which requires a live PostgreSQL/Keycloak/Valkey stack and is
+  therefore skipped in CI (`make up` required) -- confirmed CI's own
+  "Full test suite" run has never actually executed this assertion.
+- `test_start_analysis_run_recovers_the_a100_fork` seeded
+  `snapshot_sha256`/`configuration_sha256`/`code_revision_sha` with
+  `"t"`/`"u"`/`"v"`-repeated literals; none are valid hex characters,
+  so the very first insert failed its own `analysis_source_snapshot`
+  check constraint every time this test actually ran. Same CI-blind
+  gap as above -- fixed to valid hex placeholders matching this file's
+  existing convention.
+
+## [2.12.3] - 2026-08-18
+
+### Added
+
+- `make seed` inserts Late Demo public post (2026-01-13) so the
+  January 12 Demo Corp lineage and TEPP runs' knowledge cutoff is
+  falsifiable: Demo public post still opens; Late Demo does not.
+  The live post list still shows Late Demo. The cutoff filter
+  itself already lives on this stack (ADR 0016). TEPP honesty is
+  unchanged: accepted acks stay Failed transport evidence, not
+  Succeeded. Never invent a theta.
+
+## [2.12.2] - 2026-08-18
+
+### Fixed
+
+- Accepted TEPP transport evidence now stores **received** (transport
+  response) and **recorded** (row write) as distinct clocks when those
+  instants differ (ADR 0035 follow-up). After `make seed`, Demo Analyst
+  opens **TEPP measurement · Failed · Demo Corp** Measurement evidence
+  and sees one Received clock when seed receipt and persist share an
+  instant. A later start that persists in a later minute shows both
+  clocks. Digest recomputation is unchanged. Hidden runs stay 404.
+  Never invent a theta.
+
+## [2.12.1] - 2026-08-17
+
+### Fixed
+
+- A published TEPP **accepted** acknowledgement is stored as
+  **aggregate transport evidence** and stays Failed /
+  `tepp_completed_result_unsupported` (ADR 0035). After `make seed`,
+  Demo Analyst opens that Failed Demo Corp row to read contract
+  version, accepted run id, clocks, and a copyable SHA-256. The
+  section says completed-artifact identity is unavailable until TEPP
+  publishes a versioned completed-result contract. A
+  LineageWeave-local `time_multilevel_multi_affiliation` envelope, or
+  any other unpublished completed shape, stays Failed /
+  `tepp_result_not_persisted` and must not stamp Succeeded. Missing
+  `TEPP_TRANSPORT_URL` stays Failed / `tepp_not_available`. Never
+  invent a theta.
+
+## [2.12.0] - 2026-08-17
+
+### Added
+
+- A persistable TEPP **time / multilevel / multi-affiliation** result
+  is stored on the analysis-run and marked Succeeded (ADR 0034). After
+  `make seed`, Demo Analyst sees **TEPP measurement · Succeeded · Demo
+  Corp** next to the Failed missing-transport row. Home list and detail
+  show measured clocks and affiliation counts. A screen reader on that
+  Succeeded row hears open the run to read those aggregates, not only
+  the title. An `accepted` ack or an envelope this product cannot store
+  stays Failed / `tepp_result_not_persisted`. Missing
+  `TEPP_TRANSPORT_URL` stays Failed / `tepp_not_available`. Never
+  invent a theta.
+
+## [2.11.0] - 2026-08-17
+
+### Added
+
+- Home now shows the authorized customer-group tree (Group / Company /
+  Plant) instead of only a flat corp list. After `make seed`, Demo
+  Analyst walks Demo Group → Demo Corp → Demo Plant; a click opens that
+  entity as the corporate-entity report grouping. The post-scoped
+  affiliate tree is unchanged (ADR 0033).
+- Abbreviations on a post are cross-checked against that tree through
+  the existing Searxng client. A unique corroborated hit binds; a down,
+  empty, or tied search stays unbound and does not invent a parent or
+  AUTO row. Seeded `DC` on Demo Corp is synthetic Demo Corp only.
+
+## [2.10.4] - 2026-08-17
+
+### Fixed
+
+- After a Demo Corp lineage reconstruction has started, the same
+  granted retention purge empties `analysis_run_lineage_edge`,
+  `analysis_run_reconstruction`, and `analysis_source_snapshot_member`
+  when those tables exist, including their delete-reject triggers
+  (ADR 0032). Follow the same grant + admin + phrase path — do not
+  `DISABLE TRIGGER` as superuser.
+
+## [2.10.3] - 2026-08-17
+
+### Fixed
+
+- Opening a listed analysis-run that then 404s drops that stale row
+  from the home list after an authorized re-read. The next action is
+  announced as a status alert: open a remaining visible run, or
+  request a lineage reconstruction. The message still does not name
+  the thread or the cutoff (ADR 0014 / ADR 0018).
+
+## [2.10.2] - 2026-08-17
+
+### Fixed
+
+- Opening a post whose embedded picture uses invoice-like HTML
+  (`alt="Invoice > 1000"`, unquoted `width`, newlines in the base64)
+  now shows the picture. The raw payload no longer returns when a
+  remote-only or SVG tag is the whole body. Re-export as PNG or JPEG
+  if the type is rejected. The popup, `extract_base64_images`, and
+  `chunk_by_dom` share one raster allowlist (ADR 0031).
+
+## [2.10.1] - 2026-08-17
+
+### Fixed
+
+- Analysis-run list buttons now include the kind-specific next-action
+  sentence in the accessible name (WCAG 2.2 SC 4.1.2). Open a Failed
+  TEPP row: a screen reader hears connect the measurement service, not
+  only the run title. `aria-label` replaces button contents (ADR 0014).
+  No TEPP theta is invented.
+
+## [2.10.0] - 2026-08-17
+
+### Added
+
+- Home Rankings panel fuses visible posts through `RankWeaveClient`
+  (ADR 0030). After login with the port disabled or the library
+  missing, Demo Analyst sees **Rankings · RankWeave not available**.
+  An accepted hit lists the title; click opens that post. A hidden
+  post is omitted. Never invent a fused score or a theta.
+- Period reports now persist closest and farthest leftover
+  post–criterion pairs after the IRT main effects (Jeon leftover
+  map, ADR 0028 / 0029). After `make seed`, leftover pairs sit above
+  the member list; clicking a pair opens that post. A leftover pair
+  for a hidden post is omitted the same way a hidden member is.
+
+## [2.9.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  next action after landed cited evidence: Linked post evidence is
+  current, then read Event Lineage on that post. Home list opens do
+  not add that copy. No TEPP theta is invented. No cutoff body is
+  invented (ADR 0016).
+
+## [2.8.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now puts the
+  first cited evidence immediately under the named citation next
+  action, ahead of the chat input. Home list opens still wait for a
+  citation click. No TEPP theta is invented. No cutoff body is
+  invented (ADR 0016).
+
+## [2.7.2] - 2026-08-17
+
+### Fixed
+
+- R&R person chips now read `cataloged_person_id` from
+  `post_summary_role` (ADR 0027). Open a post whose R&R names a
+  cataloged person: the chip is a button even when Keyman extraction
+  was not run on that post. Click it to walk that person, not a later
+  same-named row. Historical backfill leaves a role unbound when two
+  same-named mentions already exist.
+
+## [2.7.1] - 2026-08-17
+
+### Fixed
+
+- `POST /api/analysis-runs` records Pending lineage only (ADR 0017).
+  TEPP and period-report kinds are 422 so this path cannot invent a
+  measurement. Open Analysis runs and wait until affiliated corps
+  load; choose a corp if you walk more than one, then click
+  **Request a lineage reconstruction**. Preview the picker in
+  Storybook (`Analysis/LineageEntityPicker`). Failed TEPP stays
+  terminal on this write.
+
+## [2.7.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  first cited source after the landed first Ask answer: Linked post,
+  then open that evidence. Home list opens do not add that copy. No
+  TEPP theta is invented. No cutoff body is invented (ADR 0016).
+
+## [2.6.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now puts the
+  first Ask answer immediately under the named seed next action, ahead
+  of the chat input. Home list opens keep that answer after the input.
+  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
+
+## [2.5.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  first Ask after landed chat: What happened between these events,
+  then read that answer. Home list opens do not add that copy. No
+  TEPP theta is invented. No cutoff body is invented (ADR 0016).
+
+## [2.4.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now puts Ask
+  about this lineage immediately under the Ask next action, ahead of
+  Affiliate tree. Home list opens keep Ask after Keyman. No TEPP theta
+  is invented. No cutoff body is invented (ADR 0016).
+
+## [2.3.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now focuses
+  the Ask heading after Priya Nair related nodes are current. Home
+  list opens do not steal that focus. No TEPP theta is invented. No
+  cutoff body is invented (ADR 0016).
+
+## [2.2.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  next action after Priya Nair related nodes are current: **Ask about
+  this lineage**. Home list opens do not add that copy. No TEPP theta
+  is invented. No cutoff body is invented (ADR 0016).
+
+## [2.1.0] - 2026-08-17
+
+### Added
+
+- Opening a title marked **Updated after cutoff** now shows the body
+  that run knew beside the live rewrite. After `make seed`, open Demo
+  public post from the Demo Corp lineage run: **Body this run knew** is
+  the January follow-up; the live body names the later delivery window.
+  `GET /api/posts/{id}?as_of=` reads `source_post_revision`. Analysis-run
+  detail stays titles and clocks. A missing revision is omitted — never
+  a fabricated cutoff sentence or a TEPP theta (ADR 0025).
+
+## [2.0.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now lands
+  Priya Nair related nodes under the first-related next action, ahead
+  of Affiliate tree. Home list opens still wait for a related click.
+  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
+
+## [1.9.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  first related node after landed Ada West related: Priya Nair, then
+  read that person. Home list opens do not add that copy. No TEPP
+  theta is invented. No cutoff body is invented (ADR 0016).
+
+## [1.8.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now lands
+  Ada West related nodes under the first-Keyman next action, ahead
+  of Affiliate tree. Home list opens still wait for a Keyman click.
+  No TEPP theta is invented. No cutoff body is invented (ADR 0016).
+
+## [1.7.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now names the
+  first Keyman after landed evaluation: Ada West, then read that
+  person. Home list opens do not add that copy. No TEPP theta is
+  invented. No cutoff body is invented (ADR 0016).
+
+## [1.6.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members now puts
+  Keyman and evaluation immediately under the Event Lineage next
+  action, ahead of Affiliate tree. Home list opens keep evaluation
+  above Event Lineage. No TEPP theta is invented. No cutoff body is
+  invented (ADR 0016).
+
+## [1.5.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members names the next
+  action after the current Event Lineage node: read Keyman and
+  evaluation. Home list opens do not add that copy. No TEPP theta is
+  invented. No cutoff body is invented (ADR 0016).
+
+## [1.4.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members marks that
+  post current in the popup Event Lineage DAG, so the focused heading
+  has a you-are-here node. The home DAG stays unmarked. No TEPP theta
+  is invented. No cutoff body is invented (ADR 0016).
+
+## [1.3.0] - 2026-08-17
+
+### Added
+
+- Opening Public post from the landed Demo Corp members focuses the
+  popup Event Lineage heading, matching the next-action copy. Home
+  post-list opens do not steal that focus. No TEPP theta is invented.
+
+## [1.2.0] - 2026-08-17
+
+### Added
+
+- Opening **Public post** from the landed Demo Corp report now names
+  the next action: that post is open from Demo Corp, so read Event
+  Lineage, Keyman, and evaluation. The opened member is current. Mean
+  θ stays on the report panel. No TEPP theta is invented. No cutoff
+  body is invented (ADR 0016).
+
+## [1.1.0] - 2026-08-17
+
+### Added
+
+- Opening **Open period report 2026-W02** now puts the Demo Corp
+  report (mean θ and member posts) immediately under the named next
+  action, ahead of Other Corp and the week strip. The Public post
+  member stays clickable. Mean θ stays on the report panel. No TEPP
+  theta is invented. No cutoff body is invented (ADR 0016).
+
+## [1.0.0] - 2026-08-17
+
+### Added
+
+- Opening **Open period report 2026-W02** now names the next action on
+  the landed Demo Corp report: read its mean θ and member posts, then
+  open a post. The focused comparison chip uses the visible
+  `Corporate entity: Demo Corp` caption and the persisted mean θ
+  (WCAG 2.5.3). Changing the week still focuses the report period
+  field. Mean θ stays on the report panel. No TEPP theta is invented.
+
+## [0.99.0] - 2026-08-17
+
+### Added
+
+- Opening **Open period report 2026-W02** when the operator is already
+  on that week lands the grouping comparison strip on Demo Corp. The
+  Demo Corp chip is current and focused. Changing the week still
+  focuses the report period field. Mean θ stays on the report panel.
+  No TEPP theta is invented.
+
+## [0.98.1] - 2026-08-17
+
+### Fixed
+
+- Authorized analysis-run list and detail queries are now complete
+  parameterized SQL literals. Semgrep no longer treats the visibility
+  predicate as string-concatenated user input. The $1 / $2 / $3 binds
+  are unchanged. No TEPP theta is invented.
+
+## [0.98.0] - 2026-08-17
+
+### Added
+
+- Opening **Open period report 2026-W02** from a corporate-entity
+  analysis run also switches Report grouping to Corporate entity and
+  marks the Demo Corp grouping current. The opened report is named
+  Demo Corp, not a UUID. Mean θ stays on the report panel. No TEPP
+  theta is invented.
+
+## [0.97.0] - 2026-08-17
+
+### Added
+
+- A Succeeded period-report analysis run now opens the scored week.
+  After `make seed`, open **Period report · Succeeded · Demo Corp**
+  and click **Open period report 2026-W02**: the report period field
+  is focused on that week. Failed rows stay closed. Mean θ stays on
+  the report panel. No TEPP theta is invented.
+
+## [0.96.0] - 2026-08-17
+
+### Added
+
+- `make seed` now records **Period report · Succeeded · Demo Corp** on
+  the shared snapshot after the calibrated report tables are written
+  (ADR 0024). Open that row to confirm the cutoff posts. Mean θ stays
+  on the period-report panel. Start stays 422. No TEPP theta is
+  invented.
+
+## [0.95.0] - 2026-08-17
+
+### Added
+
+- Analysis-run detail now lists labeled outbox delivery: Claimed then
+  Delivered (ADR 0023). After `make seed`, open the Demo Corp lineage
+  run to see those times. Stream entry ids stay off the payload. No
+  TEPP theta is invented.
+
+## [0.94.0] - 2026-08-17
+
+### Added
+
+- `POST /api/analysis-runs/{id}/start` now commits Running plus one
+  durable outbox row, wakes Valkey (`analysis-run-outbox`), then
+  delivers ThreadWeave or `tepp_client` (ADR 0023). A crash after
+  Start leaves the work item; refresh finishes it. Period-report
+  stays 422. No TEPP theta is invented.
+
+## [0.93.0] - 2026-08-17
+
+### Added
+
+- `make seed` now persists the designed A-100 fork on the Demo Corp
+  Succeeded lineage run. Open that run: the revised quote and delivery
+  question follow the pricing follow-up and are buttons. Start is
+  unchanged. No TEPP theta is invented.
+
+## [0.92.0] - 2026-08-17
+
+### Added
+
+- **Start TEPP measurement** on a Pending TEPP row submits
+  `AnalysisRunRequest` through `tepp_client` (ADR 0022). A missing
+  `TEPP_TRANSPORT_URL` or a refused URL is Failed /
+  `tepp_not_available`. An accepted envelope is Failed /
+  `tepp_result_not_persisted`. Failed stays terminal: **Request a new
+  TEPP measurement** records a new Pending run. Period-report start
+  stays 422. No TEPP theta is invented.
+
+## [0.91.0] - 2026-08-17
+
+### Added
+
+- After **Start reconstruction**, the titled A-100 edges are buttons.
+  Click the revised-quote child to open the live post; click the
+  pricing-follow-up parent to open that post. A child marked
+  **Updated after cutoff** still shows the live-body warning. The
+  popup does not invent a cutoff snapshot. No TEPP theta is invented.
+
+## [0.90.0] - 2026-08-17
+
+### Added
+
+- Opening an analysis-run title marked **Updated after cutoff** now
+  shows a popup status that the body is live, not a cutoff snapshot
+  (ADR 0016). After `make seed`, open the Demo Corp lineage run and
+  click Demo public post: the warning appears above the live body
+  and tells you to compare it with this run. Demo private post and
+  the home post list do not. The popup does not invent the earlier
+  text. No TEPP theta is invented.
+
+## [0.89.0] - 2026-08-17
+
+### Added
+
+- Analysis-run detail now compares each in-cutoff title's live
+  `updated_at` with that run's knowledge cutoff. After `make seed`,
+  open the Demo Corp lineage run: Demo public post is marked
+  **Updated after cutoff**; Demo private post is not. Opening a
+  marked title still shows the live body -- cutoff body versioning
+  stays a later slice (ADR 0016). The list stays aggregates-only.
+  No TEPP theta is invented.
+
+## [0.88.0] - 2026-08-16
+
+### Added
+
+- `POST /api/analysis-runs/{id}/start` runs ThreadWeave on a visible
+  Pending lineage cutoff bag and persists run-scoped parent choices
+  (ADR 0021). Open the Pending Demo Corp row, then start reconstruction.
+  The designed A-100 fork (revised quote and delivery question under the
+  pricing follow-up) is the acceptance tree. TEPP and period-report
+  start are 422 — this path does not invent a theta. A Succeeded retry
+  returns the stored digest. A Running restart is 409. Create freezes
+  authorized post ids so start cannot pick up a later backfill. Live
+  Event Lineage stays a separate rebuild.
+
+## [0.87.0] - 2026-08-16
+
+### Added
+
+- Operators can empty a run-bearing analysis-run registry without a
+  superuser trigger disable. Insert an unrevoked
+  `analysis_run_retention_grant` for `session_user`, grant
+  `analysis_run_retention_admin`, then
+  `select purge_analysis_run_registry('approved-retention-purge')`
+  (ADR 0020). Export `analysis_run_retention_event`, delete those
+  rows, then roll back 0020 and 0018. A raw `DELETE`, a published
+  token without a grant, and a runtime role that is not the admin
+  role still fail.
+- Repeated citation chips and close buttons use named design tokens
+  in `frontend/src/styles/tokens.css`. Preview them in Storybook
+  (`cd frontend && pnpm run storybook`).
+
+## [0.86.2] - 2026-08-16
+
+### Fixed
+
+- An R&R organization button now walks the catalog id stored on that
+  role row (ADR 0019). Two catalog orgs can share a display name; open
+  the post, click the name, and you stay on the resolved org — not a
+  homonym. `GET /api/teams/{id}/related` matches person/entity authz:
+  another corp's private-only team is 403; an unknown UUID is 404.
+
+## [0.86.1] - 2026-08-16
+
+### Changed
+
+- Opening a post or its evidence panel now shows each embedded
+  `data:image` picture in document order, with the surrounding sentences
+  as text. The raw base64 string is no longer dumped into the popup.
+  Remote `http(s)` image URLs stay unloaded. After `make seed`, a post
+  whose body includes a data-URI image shows the picture; Extract Keyman
+  or Ask still runs OCR on that image for search.
+
+### Fixed
+
+- Opening a Pending lineage run repeats that reconstruction has not
+  started. Pending next-action copy is pinned to the registered run
+  kinds, so a Pending TEPP row does not say reconstruction.
+
+## [0.86.0] - 2026-08-16
+
+### Added
+
+- Related-node walks now include team and organization mention edges.
+  After `make seed` and a summary that names 설계팀 on two posts, open
+  either post, click the R&R team, and open the sibling post (ADR 0018).
+  A team-only follow-up is no longer an island.
+- `GET /api/teams/{team_id}/related` starts the same RWR walk Keyman
+  and corporate-entity related already use. Related team chips are
+  buttons.
+
+### Fixed
+
+- Thread-group analysis-run *lists* now require an in-cutoff visible
+  post. A later public post in that thread group no longer surfaces a
+  January run the account was not allowed to know.
+- Failed period-report rows tell the operator to rebuild the report.
+  Next-action copy is pinned to the registered run kinds. A pending
+  TEPP corpus does not claim a calibrated measurement.
+
+## [0.85.0] - 2026-08-16
+
+### Added
+
+- `POST /api/analysis-runs` records a Pending lineage or TEPP run on an
+  authorized cutoff capture (ADR 0017). The home panel's **Request a
+  lineage reconstruction** button writes that row so an operator can
+  confirm the cutoff corpus immediately. Reconstruction and live TEPP
+  execution stay later slices — this write never invents a theta.
+- Failed lineage rows tell the operator to retry reconstruction; only
+  Failed TEPP rows mention the measurement service. Pending rows say
+  reconstruction has not started yet.
+
+## [0.84.1] - 2026-08-16
+
+### Fixed
+
+- Analysis-run detail keeps 12-character digest prefixes as visible
+  text (so assistive technology hears `Code` / `Config` values) and
+  puts the full digest on hover. Open the Demo Corp lineage run, hover
+  a prefix, and match it to the API payload. The home list still hides
+  digests even when the list JSON includes them.
+- Opening a cutoff title now says the live body may have changed after
+  that run. Compare the opened post with the cutoff date before you
+  treat it as reconstructed evidence (ADR 0016).
+
+## [0.84.0] - 2026-08-16
+
+### Added
+
+- `make seed` records a Demo Corp TEPP measurement run through
+  `tepp_client` on the same snapshot as the lineage run (ADR 0013).
+  The default transport is unavailable, so the home list shows
+  "TEPP measurement · Failed · Demo Corp" and tells the operator to
+  open the run, then connect the measurement service. Detail history
+  keeps `tepp_not_available` -- never a fabricated theta. TEPP stays
+  a wire client, not a local psychometric engine. `make seed` skips
+  snapshot-count inserts once counts exist so a re-run does not hit
+  the freeze trigger. A failed lineage row tells the operator to retry
+  reconstruction; only a failed TEPP row mentions the measurement
+  service. A failed period-report row tells the operator to rebuild
+  the report from a current snapshot. A pending TEPP row does not
+  claim a calibrated measurement. Stacked PRs now run the same
+  GitHub Checks as PRs to main.
+
+## [0.83.0] - 2026-08-16
+
+### Fixed
+
+- Analysis-run detail now lists only ABAC-visible posts whose
+  `created_at` is at or before that run's `knowledge_cutoff`. After
+  `make seed`, open the Demo Corp lineage run: Demo public post is
+  there; a later own-corp follow-up is not. The live post list is
+  unchanged. Click a listed title to inspect what that cutoff
+  reconstructed (ADR 0016).
+- Upgrading through `0016_cross_post_actor_identity.sql` copies R&R
+  person names into `post_summary_person_mention` and leaves Keyman
+  `post_person_mention.mention_context` in place. Re-run Keyman only
+  when you want a new Keyman set -- a later summary no longer erases
+  the stolen row.
+
+## [0.82.0] - 2026-08-16
+
+### Added
+
+- Analysis-run detail lists ABAC-visible posts in the run's scope.
+  After `make seed`, the Demo Corp lineage run opens the Demo public
+  post. Hidden other-corp private posts never appear. List payloads
+  stay aggregates-only.
+
+## [0.81.0] - 2026-08-16
+
+### Added
+
+- Analysis-run detail shows the labeled lifecycle: Pending, Running,
+  then Succeeded, with occurrence times from `analysis_run_status_event`.
+  The list stays latest-status only. Hidden runs still 404 and never
+  leak events. Failure codes stay machine tokens -- no invented label.
+  Synthetic Demo Corp seed only.
+
+## [0.80.0] - 2026-08-16
+
+### Added
+
+- Home Analysis runs rows are buttons. Clicking the seeded Demo Corp
+  lineage run opens `GET /api/analysis-runs/{id}` and shows cutoff,
+  requested date, and document count. A hidden run is "This analysis
+  run is not visible." -- never a raw 404 or a DSN. Still synthetic
+  aggregates only.
+
+## [0.79.0] - 2026-08-16
+
+### Added
+
+- Authorized analysis-run evidence on the product home page. After
+  `make seed`, Demo Analyst sees "Lineage reconstruction · Succeeded ·
+  Demo Corp" with the synthetic document count. `GET /api/analysis-runs`
+  is scoped in SQL: another tenant's run 404s and never appears in the
+  list. The payload is labels and aggregates -- never source SQL, a DSN,
+  or a raw record. TEPP stays behind `tepp_client`; Null channels are
+  unchanged.
+
+## [0.78.0] - 2026-08-15
+
+### Changed
+
+- Related-node person chips now use the localized `person_side` lookup label
+  supplied by the authorized API payload. Users see business context such as
+  `Our side` or `Counterparty`, while ontology class metadata remains available
+  separately for semantic processing and provenance.
+- Related-person buttons now expose that same caption in the accessible name, so
+  assistive technology hears `Related nodes for Priya Nair (Counterparty)`
+  instead of the name alone.
+- Structured extraction, summarization, commitment, relationship-classification,
+  and LLM-as-a-Judge consumers now request contextual-orchestrator `auto` mode
+  so the orchestration plane can meet the quality requirement and then minimize
+  known execution cost. Explicit checked `verify` paths remain unchanged
+  (ADR 0015).
+
+## [0.77.0] - 2026-08-14
+
+### Fixed
+
+- Keyman and R&R person mentions now replace independent source projections. Knowledge Graph edges have one canonical identity plus post-level evidence, so removed actors and concurrent writes cannot leave stale or duplicate buyer-visible relationships.
+- Vision-response parsing now strips balanced outer Markdown emphasis from field values
+  while still accepting emphasized field labels, so OCR such as
+  ``TEXT: **LT7**`` is not truncated.
+- A real live synthetic regression batch run surfaced a genuine
+  `DeadlockDetectedError` from concurrent corporate-entity creation:
+  two concurrent transactions each creating a different new entity,
+  mentioned in opposite order across two different posts, took
+  row-level locks in opposite order and deadlocked. Entity *creation*
+  (the rare, first-mention-only branch) now serializes through a
+  single named Postgres advisory transaction lock, taken only right
+  before the write and auto-released at commit/rollback -- the
+  lock-free similarity-matching fast path every already-cataloged
+  entity resolves through is unaffected. See ADR 0012.
+
+## [0.76.0] - 2026-08-14
+
+### Added
+
+- Standards-complete W3C PROV-O support: all 30 classes, all 50
+  normative properties, both qualification tables, qualified-to-
+  unqualified implication, property hierarchy, defined inverses,
+  Appendix B inverse-name normalization, RDF serialization, and a
+  normalized PostgreSQL assertion store with fail-closed domain,
+  range, object-kind, and datatype enforcement (ADR 0011).
+- A dedicated exact-head PROV-O contract workflow runs the complete
+  registry/inference suite, real PostgreSQL migration tests, public
+  docstring checks, and 100% statement/branch coverage for the owned
+  runtime module.
+
+### Changed
+
+- The product navigation graph remains an explicit projection;
+  literal-valued and qualified provenance is no longer forced into
+  `knowledge_graph_edge`.
+
+### Fixed
+
+- Review hardening verifies complete hierarchy placement, rejects parent
+  failures and cycles, propagates canonical affiliations, replaces stale
+  actor projections, enforces atomic team identity, validates timezone-aware
+  `xsd:dateTime` literals (including the XSD `±14:00` offset bound), and
+  protects referenced provenance rows.
+
+## [0.75.0] - 2026-08-14
+
+### Added
+
+- A real counterparty organization mentioned for the first time now
+  gets auto-created into the corporate hierarchy, not left permanently
+  unresolved. Synthetic regression fixtures prove the first-mention gap.
+  An LLM proposes a
+  Group/Company/Plant placement from context; a real new
+  `corporate_entity` row is only created once the proposal is
+  search-corroborated (reusing the existing Searxng verification
+  client, no new search integration). Auto-created rows get a
+  deterministic `AUTO-`-prefixed code, kept structurally separate from
+  the real login corp-code namespace. Wired into both Keyman
+  affiliation resolution and R&R organization-actor resolution.
+
+## [0.74.0] - 2026-08-14
+
+### Added
+
+- R&R team and organization actors now get a shared identity across
+  posts, not just per-post free text -- the same "설계팀" (design
+  team) named in ten posts resolves to one `cataloged_team` row
+  (identity key: team name + parent org, since a bare team name is not
+  by itself identifying), and an organization actor resolves against
+  the existing `corporate_entity` catalog. Each resolved actor gets a
+  real Knowledge Graph mention edge (`edge_mention_team`,
+  `edge_team_affiliation`, `edge_mention_organization`), so extraction
+  results now genuinely become cross-post lineage clues instead of
+  per-post islands. A person R&R actor is opportunistically joined to
+  an existing Keyman-cataloged person by name (documented gap: R&R does
+  not yet originate new person identities itself -- see ADR 0009).
+
+## [0.73.0] - 2026-08-14
+
+### Fixed
+
+- Image OCR/caption parsing no longer discards a real vision response
+  just because its formatting was close but not exact (bolded labels
+  like `**TEXT:**`, reordered labels, or a missing TAGS line) --
+  observed live against synthetic embedded-image fixtures in the synthetic regression batch
+  in format-variation fixtures. Fields are now recovered independently; only a
+  response with neither TEXT nor CAPTION content is treated as
+  unusable. A strict format mismatch was silently producing the same
+  "[image: content unavailable]" placeholder as a genuinely unavailable
+  vision channel, discarding real, already-paid-for content.
+
+## [0.72.0] - 2026-08-14
+
+### Added
+
+- Abbreviated/slang organization names (e.g. "AGP") are now resolved
+  to their canonical name ("Aurora Grid Power") via LLM context, then
+  cross-verified against external search before being trusted -- new
+  `lineageweave/organization_name_resolution.py`, reusing the existing
+  Searxng verification client rather than a second web-search
+  integration. Cached in a new `organization_name_resolution` table
+  keyed by the raw name.
+- Wired into Keyman affiliation ingestion (both the API path and the
+  offline synthetic-batch script): a search-corroborated resolution feeds
+  `resolve_corporate_entity`, an unverified one leaves the raw name
+  unchanged.
+
+### Fixed
+
+- The offline synthetic-batch script's own re-implementation of Keyman
+  affiliation persistence was missing `role_title` entirely (a stale
+  copy that predated that feature) -- fixed alongside this change.
+
+## [0.71.0] - 2026-08-14
+
+### Added
+
+- VOC evidence now quotes the extractive excerpt under the counterparty
+  it names. After `make seed`, Northridge Grid in VOC evidence shows
+  the sentence that mentions it, instead of a detached excerpt list
+  above the names. Unassigned excerpts stay in the list; a post with
+  no named organization still says so.
+
+## [0.70.0] - 2026-08-14
+
+### Added
+
+- R&R's `actor_type_code` gains `prov_team`, a meso-level actor type for
+  a named sub-unit of a company (e.g. "설계팀"/design team) -- distinct
+  from both a person and the company itself. Grounded in the W3C
+  Organization Ontology's `org:OrganizationalUnit` (Reynolds, 2014).
+- A team actor now requires `affiliated_organization_name` in the same
+  way a person actor does -- a team's own name never answers "which
+  company."
+- R&R badge shows a distinct "Team" label/color, not the prior binary
+  Person/Organization ternary (which would have mislabeled a team as
+  "Organization").
+
+## [0.69.0] - 2026-08-14
+
+### Added
+
+- Keyman extraction now captures a stated job title/position
+  (`PersonMention.job_title`), since two different real people can
+  share a name and a title is real evidence for telling them apart.
+  Persisted to `person_affiliation.role_title` (an existing schema
+  column, previously never populated) and a new
+  `cataloged_person.last_known_job_title` for a title stated without a
+  named organization to attach it to.
+- `_upsert_person` no longer blindly merges a same-name+side match: a
+  genuinely conflicting stated title creates a fresh person row instead
+  of reusing one, verified by a real test with two posts naming the
+  same name and different titles.
+- Keyman panel shows the person's title next to their name. After
+  `make seed`, Ada West is "Account manager" and Priya Nair is
+  "Procurement lead" so the title is visible without a live LLM.
+
+### Fixed
+
+- `scripts/seed_demo_data.py` no longer embeds the local Keycloak admin
+  password. `make seed` still supplies the compose default via
+  `KEYCLOAK_ADMIN_PASSWORD`; a direct script run requires that env var
+  or `--keycloak-admin-password`.
+
+## [0.68.0] - 2026-08-14
+
+### Changed
+
+- R&R's named actor is no longer forced into a person slot. A
+  business post can name an organization acting in its own name
+  ("당사," "Demo Corp"), not an individual --
+  `RoleResponsibility.actor_name` (renamed from `person_name`) now
+  carries `actor_type_code` (Person / Organization, W3C PROV-O
+  grounded) and an LLM-inferred `affiliated_organization_name` for
+  person actors. The popup's R&R list shows a Person/Organization
+  badge and the inferred affiliation; only a person actor still links
+  to the Keyman panel. See ADR 0006.
+
+## [0.67.0] - 2026-08-14
+
+### Added
+
+- Keyman affiliation names that resolve to a cataloged org start
+  the same related-node walk as a related-entity click. After
+  `make seed`, Demo Corp next to Ada West opens
+  `GET /api/corporate-entities/{id}/related`. Unresolved affiliation
+  names stay text -- never a guessed neighborhood.
+
+## [0.66.0] - 2026-08-14
+
+### Added
+
+- Classified counterparty names that resolve to a cataloged org start
+  the same related-node walk as an affiliate-tree org click. After
+  `make seed`, Demo Corp in Counterparties opens
+  `GET /api/corporate-entities/{id}/related`. Unresolved names
+  (Northridge Grid) stay text -- never a guessed neighborhood.
+
+## [0.65.0] - 2026-08-14
+
+### Added
+
+- Resolved affiliate-tree organizations start the same related-node
+  walk as a Keyman org click. After `make seed`, Demo Corp / Test
+  Corp in the tree open `GET /api/corporate-entities/{id}/related`.
+  Unresolved names stay text -- never a guessed neighborhood.
+
+## [0.64.0] - 2026-08-14
+
+### Added
+
+- Related corporate-entity nodes in the Keyman walk are the same
+  RWR as a person. Clicking `Demo Corp` / `Test Corp` loads
+  `GET /api/corporate-entities/{id}/related` so the buyer can
+  continue the walk from the org. An entity with no visible
+  affiliation stays 403 -- never a guessed neighborhood.
+
+## [0.63.0] - 2026-08-14
+
+### Added
+
+- A third seeded Ask question, "What is the next commitment?", on
+  every fixture that already answers the first two chips. A-100
+  names Send Northridge Grid the revised quote due 2026-01-12;
+  B-200 names the Westfield specification due 2026-01-14; Riverbend
+  names its calendar ticket due 2026-01-09; rec-006 says it has no
+  open commitment. After `make seed` the popup chips connect Ask
+  to the same dated tickets home Calendar already lists.
+
+## [0.62.0] - 2026-08-14
+
+### Added
+
+- VOC evidence counterparties that already appear on the affiliate
+  tree are the same Keyman walk. Clicking `Northridge Grid` loads
+  related nodes for Priya Nair so the buyer does not have to find
+  the same org again under Affiliate tree.
+
+### Security
+
+- Verification badges only become links for `http:` / `https:`
+  evidence URLs. A `javascript:` or `data:` value stays a non-link
+  badge.
+
+## [0.61.0] - 2026-08-14
+
+### Changed
+
+- Activity event-type badges show Ticket created / Status changed /
+  Commitment derived instead of snake_case codes. Unknown types stay
+  as the raw code; the badge never invents a name.
+
+## [0.60.0] - 2026-08-14
+
+### Changed
+
+- Activity-feed `ticket_status_changed` summaries use the lookup
+  label (`In progress`, `Closed`) instead of the raw code. Same
+  `common_lookup_value` labels the popup select, Calendar, and
+  period-report members already show.
+
+## [0.59.0] - 2026-08-14
+
+### Added
+
+- Period-report members show the ticket status lookup label next to
+  the title and due date. After `make seed` the A-100 pricing
+  follow-up reads Open plus `due 2026-01-12`, same label the
+  Calendar and popup select already use.
+
+## [0.58.0] - 2026-08-14
+
+### Changed
+
+- After an unmatched Verify 503, the Verify against web search
+  button is hidden the same way as Evaluate / Extract / Derive.
+  Seeded counterparty badges stay; the button that can only 503
+  again does not. Still never fabricates a search result.
+
+## [0.57.0] - 2026-08-14
+
+### Added
+
+- Home Calendar rows show the ticket status lookup label next to the
+  due date. After `make seed` the A-100 pricing ticket reads Open
+  plus `due 2026-01-12`, same label the popup select already uses.
+
+## [0.56.0] - 2026-08-14
+
+### Added
+
+- A second seeded Ask question, "Who is involved?", on every fixture
+  that already answers "What happened between these events?" A-100
+  names Ada West and Priya Nair; B-200 names Jordan Hale; rec-006
+  and Riverbend say they do not name a Keyman. After `make seed`
+  the popup shows two chips instead of one canned prompt.
+
+## [0.55.0] - 2026-08-14
+
+### Added
+
+- Ticket status on the popup select shows `common_lookup_value`
+  labels (`Open`, `In progress`, `Closed`) instead of raw codes.
+  Codes stay on the payload. A missing lookup falls back to the code.
+
+## [0.54.0] - 2026-08-14
+
+### Changed
+
+- After an unmatched 503, Evaluate, Extract Keymen, and Derive
+  commitment are hidden the same way as free-text Ask. Seeded
+  evaluation, Keyman, and ticket rows stay; the buttons that can
+  only 503 again do not. Still never fabricates a live LLM result.
+
+## [0.53.0] - 2026-08-14
+
+### Added
+
+- `make seed` opens a ticket on the B-200 specification-revision
+  fixture. A B-200 report-member click now shows "Send Westfield
+  Power the revised specification" (due 2026-01-14) instead of
+  "No tickets yet," and home Calendar lists it next to the A-100
+  pricing ticket.
+
+## [0.52.0] - 2026-08-14
+
+### Changed
+
+- After an unmatched Ask 503, the free-text input is hidden. Seeded
+  answers stay on screen with "Only seeded questions can be answered
+  without an orchestrator." Still never fabricates a live reply.
+
+## [0.51.0] - 2026-08-14
+
+### Added
+
+- Home Calendar lists the seeded A-100 pricing ticket due 2026-01-12.
+  After `make seed`, GET /api/calendar includes "Send Northridge Grid
+  the revised quote" next to the Riverbend commitment, so the buyer
+  sees the same dated ticket already shown on the period-report row.
+
+## [0.50.0] - 2026-08-14
+
+### Added
+
+- Seeded fixture summaries include R&R from the Event Lineage cast.
+  An A-100 DAG click now shows Ada West / Priya Nair, and a B-200
+  click shows Jordan Hale -- not an empty R&R list. rec-006 and the
+  calendar commitment stay empty because they have no cast people.
+  An R&R name that matches a Keyman opens the same related-node walk
+  as the Keyman and affiliate-tree buttons.
+
+## [0.49.0] - 2026-08-14
+
+### Added
+
+- Period-report members show the earliest open ticket title next to
+  the due date. The home-page report list now reads "Send Northridge
+  Grid the revised quote" plus `due 2026-01-12` on the A-100 pricing
+  follow-up, so the buyer sees the seeded ticket before opening the
+  post.
+
+## [0.48.0] - 2026-08-14
+
+### Added
+
+- Period-report members include the earliest open ticket due date.
+  The home-page report list shows `due 2026-01-12` on the A-100
+  pricing follow-up so the buyer sees the seeded ticket before
+  opening the post.
+
+## [0.47.0] - 2026-08-14
+
+### Added
+
+- Seeded Ask answers for the demo public post, A-100/B-200 Event
+  Lineage fixtures, and the Riverbend calendar post. `GET`/`POST
+  /api/posts/{id}/chat` read `post_chat_result` first so the popup
+  transcript answers "What happened between these events?" without an
+  orchestrator. An unmatched question still 503s -- never a fabricated
+  live reply.
+
+## [0.46.0] - 2026-08-14
+
+### Changed
+
+- Evaluate, Extract Keymen, Derive commitment, and Verify show the
+  same kind of empty state as chat when the backing service is 503
+  (`Evaluation unavailable…`, `Keyman extraction unavailable…`,
+  `Commitment derivation unavailable…`, `Verification unavailable…`)
+  instead of `HTTP 503`. Still never fabricates a result.
+
+## [0.45.0] - 2026-08-14
+
+### Changed
+
+- Chat 503 (orchestrator unset) shows
+  `Chat unavailable (LLM orchestrator not configured).` instead of
+  `HTTP 503`. Still never fabricates an answer.
+
+## [0.44.0] - 2026-08-14
+
+### Added
+
+- `make seed` XADDs a `ticket_created` event onto Valkey for the
+  seeded A-100 and calendar tickets. A report-member click now shows
+  "Ticket created: Send Northridge Grid the revised quote" instead of
+  "No activity yet."
+
+## [0.43.0] - 2026-08-14
+
+### Added
+
+- VOC evidence counterparties show the same verification badge as the
+  Counterparties panel (`Not yet checked` / `Corroborated` plus the
+  evidence URL when present). The extractive quote is not an unchecked
+  second claim.
+
+## [0.42.0] - 2026-08-14
+
+### Added
+
+- `make seed` opens tickets on the A-100 pricing follow-up and
+  delivery-schedule fixtures. A report-member click now shows
+  "Send Northridge Grid the revised quote" (due 2026-01-12) instead
+  of "No tickets yet."
+
+## [0.41.0] - 2026-08-14
+
+### Changed
+
+- Period-report members with Event Lineage, Keyman, and evaluation
+  sort ahead of dummy high/low band rows. A-100 fixtures score on the
+  high process unit and B-200 on the low unit, so a member click opens
+  a reconstruct post that already has those panels.
+
+## [0.40.0] - 2026-08-14
+
+### Added
+
+- Seeded period reports fold A-100/B-200 Event Lineage fixtures and
+  the Riverbend calendar post (already carrying IRT cells) into the
+  shared-metric 2026-W02 bank. The comparison strip can open those DAG
+  posts; dummy high/low band posts remain for the PU ranking test.
+
+## [0.39.0] - 2026-08-14
+
+### Added
+
+- Affiliate-tree people are the same Keyman walk as the Keyman panel.
+  Clicking `Ada West` or `Priya Nair` on the tree loads related nodes
+  (`GET /api/keymen/{id}/related`) so the buyer does not have to find
+  the same name again under Keyman.
+
+## [0.38.0] - 2026-08-14
+
+### Added
+
+- Seeded Keyman and VOC counterparty for B-200 Event Lineage fixtures
+  (Jordan Hale / Westfield Power, Voice of Market). Clicking a B-200
+  DAG node is no longer an empty Keyman or VOC panel. A-100 and
+  rec-006 casts are unchanged.
+
+## [0.37.0] - 2026-08-14
+
+### Added
+
+- Seeded Keymen, affiliate-tree leaves, and VOC counterparties for
+  every A-100 proj-alpha Event Lineage fixture and the calendar
+  Riverbend post. Clicking a DAG node now shows Ada West / Priya Nair,
+  Northridge Grid, and an extractive excerpt without a live
+  orchestrator. rec-006 stays uncast so it remains its own root. A
+  post with no named org still has empty VOC excerpts -- never a
+  fabricated quote.
+
+## [0.36.0] - 2026-08-14
+
+### Added
+
+- `make seed` writes constructed IRT rubric cells for the demo public
+  post, A-100/B-200 Event Lineage fixtures, and the calendar
+  commitment. The Post quality panel shows those categories after a
+  fresh stack -- not "Not yet evaluated." Cells are title-derived, not
+  an LLM judge; thetas still come only from `calibrate_period_report`.
+
+## [0.35.0] - 2026-08-13
+
+### Added
+
+- Post list and popup show `common_lookup_value` labels for VOC type
+  and visibility (`Voice of Customer`, `Public`) instead of raw codes.
+  Codes stay on the payload. A missing lookup falls back to the code.
+
+## [0.34.0] - 2026-08-13
+
+### Added
+
+- Seeded Korean summaries for every A-100/B-200 reconstruct fixture and
+  the calendar commitment post. Clicking an Event Lineage DAG node or a
+  calendar row now serves `GET /api/posts/{id}/summary` from
+  `post_summary_result` without a live orchestrator. A non-fixture post
+  with no stored row still 503s -- never a fabricated summary.
+
+## [0.33.0] - 2026-08-13
+
+### Added
+
+- Home-page grouping comparison strip: `GET /api/reports/compare/{period}`
+  returns every process unit, corporate entity, and thread group scored
+  on the shared FIPC metric. Rebuild now writes all three kinds.
+  Clicking a row switches the Period reports grouping. Seed assigns
+  high posts to thread A-100 and low posts to B-200 so the strip is
+  not empty after `make seed`.
+
+## [0.32.0] - 2026-08-13
+
+### Added
+
+- CAT item-information selection on the shared FIPC bank: after a
+  group is scored, `information_polytomous` ranks rubric items by
+  Fisher information at that group's mean θ (Lord, 1980). Rank 1 is
+  the max-info pick -- a high PU and a low PU can be measured by
+  different items on the same bank. Rankings persist to
+  `report_item_information` and the Period reports panel labels
+  `CAT: sales-lead I=0.70`. Information comes from fast-mlsirm's
+  Rust GRM/GPCM curves, never a hand-rolled I(θ).
+
+## [0.31.0] - 2026-08-13
+
+### Added
+
+- Shared-metric FIPC: the first period free-calibrates one item bank
+  on the pooled posts; every process unit (and other grouping) is
+  then scored on that bank so a high PU and a low PU stay comparable.
+  Independent per-group refits would each re-center near 0. The
+  Period reports panel labels first-period groups `shared metric`.
+
+## [0.30.0] - 2026-08-13
+
+### Added
+
+- Persisted Korean summaries (`post_summary_result` /
+  `post_summary_event` / `post_summary_role`). `GET /api/posts/{id}/summary`
+  returns a stored row first so a seeded demo popup is not empty when
+  the LLM orchestrator is off. Live derivation still writes through
+  the same tables. Seed writes a synthetic summary for the demo public
+  post. A missing stored row and a missing LLM stay 503 -- never a
+  fabricated summary.
+
+## [0.29.0] - 2026-08-13
+
+### Added
+
+- True cross-period FIPC linking: the first week for a grouping
+  free-calibrates and persists `report_item_parameter`; later weeks
+  EAP-score on those fixed item parameters so mean θ is comparable
+  across weeks (Kim, 2006). An independent all-high week would
+  re-center near 0; the linked week stays high on the reference
+  metric. `GET /api/reports/{grouping}` lists the trend; the Period
+  reports panel shows `vs 2026-W02: +0.92`. Seed writes a constructed
+  2026-W02 reference and 2026-W03 all-high follow-up.
+
+## [0.28.2] - 2026-08-13
+
+### Added
+
+- Period reports lists each member post with its fitted EAP θ and
+  opens that post on click -- the same pattern as the calendar.
+
+## [0.28.1] - 2026-08-13
+
+### Added
+
+- `make seed` inserts constructed high/low IRT rows for 2026-W02 and
+  persists a real FIPC report, so the Period reports panel is not empty
+  after a fresh stack. Thetas still come only from
+  `calibrate_period_report`.
+
+## [0.28.0] - 2026-08-13
+
+### Added
+
+- ADR 0003 slice 3: `lineageweave/period_report.py` fits GRM and GPCM
+  with `fast_mlsirm.fit_polytomous`, EAP-scores posts, and picks the
+  model via `fixed_item_calibration_diagnostics`. Scores persist to
+  `report_period_score` / `report_member_score`.
+  `GET /api/reports/{grouping}/{period}` and `POST .../rebuild`
+  (`post_admin`) plus a home-page Period reports panel. The accuracy
+  test constructs high vs low category groups and asserts the fitted
+  mean θ ranks them correctly -- no hardcoded score.
+
+## [0.27.0] - 2026-08-13
+
+### Added
+
+- ADR 0003 slice 2: a post can be LLM-as-a-Judge evaluated through
+  `fast-mlsirm`'s `ContextualOrchestratorJudge`. Scores persist only
+  via `LLMJudgeResult.to_irt_row()` into `post_evaluation_response`
+  (one row per post per rubric criterion). Null client is unavailable,
+  never a fabricated score.
+- Versioned rubric: constructive stance, negative stance, sales-lead
+  specificity. `POST /api/posts/{id}/evaluate` +
+  `GET /api/posts/{id}/evaluation`, and an "Evaluate post" panel on
+  the post popup.
+
+## [0.26.0] - 2026-08-13
+
+### Added
+
+- External search verification for Knowledge Graph relation inferences:
+  `lineageweave/relation_verification.py`'s `SearxngRelationVerificationClient`
+  checks an LLM-classified counterparty organization/relationship
+  against a self-hosted Searxng instance, catching a hallucinated
+  organization name with zero real-world footprint (FEVER grounding:
+  Thorne, Vlachos, Christodoulopoulos, & Mittal, 2018). A search hit
+  only corroborates when a distinctive org-name token appears in the
+  result host or snippet and the host is not itself a search page --
+  engines echo any query in the title, so "any result" is not evidence.
+- New Docker Compose service `searxng` (`docker/searxng/`), wired into
+  the backend via `SEARXNG_BASE_URL`.
+- `post_counterparty_entity` gained `verification_status_code`
+  (`verify_pending` / `verify_corroborated` / `verify_uncorroborated`),
+  `verification_evidence_url`, `verification_checked_at`
+  (`migrations/0004_relation_verification.sql`). A re-classification
+  resets these back to pending.
+- New `POST /api/posts/{id}/verify-relations` endpoint and a
+  `CounterpartyPanel` UI section (status badge, linked to the evidence
+  URL when corroborated, with a "Verify against web search" action).
+- ADR 0005 documents the design decision.
+
+## [0.25.0] - 2026-08-13
+
+### Added
+
+- Post bodies are normalized before ever reaching an LLM or embedding
+  call: HTML tags are stripped to clean text, per-block formatting
+  (`style`, heading level) is kept as separate `FormattingHint` metadata
+  instead of being embedded, and base64-embedded images are described
+  by a vision-capable model and placed as `[image: caption | text: ocr]`
+  at their original document position instead of being sent raw or
+  dropped (`lineageweave/post_content_normalization.py`, VIPS grounding:
+  Cai, Yu, Wen, & Ma, 2003). Plain-text comparison operators (`qty < 50`)
+  are not treated as HTML.
+- `chunking.py`'s DOM chunker now captures each block's `style`
+  attribute and splits on `h1`-`h6` heading boundaries.
+- New `VISION_MODEL` setting; `extract-keymen`, post summary, commitment
+  derivation, and chat source retrieval (including linked posts pulled
+  in as RAG context) all normalize `post_body` before the LLM sees it.
+  The raw post-detail read (`GET /api/posts/{id}`) is intentionally
+  left untouched so the frontend still renders the post as-authored.
+
+## [0.24.0] - 2026-08-13
+
+### Added
+
+- Related Keyman nodes are now walkable: a related post opens that
+  post, a related person loads *their* RWR neighbourhood. Dead text
+  next to "Related to Ada West" is no longer a dead end.
+- Chat citation chips show the source post's title
+  (`cited_posts` from `POST /api/posts/{id}/chat`) instead of a
+  truncated UUID. `cited_post_summaries` drops unknown ids rather than
+  inventing a label.
+
+## [0.23.0] - 2026-08-13
+
+### Added
+
+- Keyman side and corporate-entity level now carry lookup labels from
+  `common_lookup_value` (`person_side_label`, `entity_level_label`).
+  The popup shows "Ada West (Our side)" and "Demo Corp (Company)"
+  instead of raw `our_side` / `company` / `plant` codes. Missing lookup
+  rows still fall back to the code, never an invented name.
+
+## [0.22.0] - 2026-08-13
+
+### Added
+
+- Related Keyman nodes now carry the ontology class (`ontology_iri`,
+  `ontology_label`) from `docs/ontology/lineageweave-kg.ttl`, so the
+  popup shows "Priya Nair (Person)" instead of the raw `node_person`
+  lookup code. Missing terms stay unlabeled rather than inventing a
+  string -- same missing-vs-negative rule as every Null channel.
+  `GET /api/keymen/{id}/related` is the wire; the frontend falls back
+  to `node_type_code` only when the ontology has no term.
+
+### Fixed
+
+- `frontend/package.json` was invalid JSON after #24 dropped the
+  comma on the version line, so `pnpm run test` / CI frontend could
+  not even parse the manifest.
+
+## [0.21.2] - 2026-08-13
+
+### Added
+
+- The Knowledge Graph is now grounded in a real, machine-checked
+  Ontology and Semantic Layer (see
+  [ADR 0004](docs/adr/0004-knowledge-graph-ontology.md)) -- the
+  brief's latest revision required this everywhere the KG is used
+  (Keyman traversal, corporate hierarchy tree, entity-relationship
+  classification, indirect lineage linking, in-popup chat evidence).
+  `docs/ontology/lineageweave-kg.ttl`: a real OWL 2 / RDFS / SKOS
+  ontology (Turtle) formalizing `knowledge_graph_edge`'s node/edge
+  vocabulary and the `entity_relationship_type`/`person_side`/
+  `corporate_entity_level` controlled vocabularies -- classes with
+  declared subclass relations, object properties with
+  `rdfs:domain`/`rdfs:range`, and the corporate-level hierarchy
+  (Group -> Company -> Plant) as a proper SKOS concept scheme.
+  `lineageweave/ontology.py` parses it with `rdflib` and exposes the
+  vocabulary as importable IRI constants.
+- `tests/test_ontology.py`: a real round-trip check, not just "does
+  the file parse" -- every lookup code `scripts/seed_demo_data.py`
+  actually seeds (for the categories this ontology covers) must have
+  a matching ontology term, and vice versa, so the ontology cannot
+  silently drift from the relational schema's real vocabulary.
+
+## [0.21.1] - 2026-08-13
+
+### Fixed
+
+- `test_seed_calendar_commitment_surfaces_on_get_calendar` drives the
+  same `_seed_demo_calendar_commitment` helper `make seed` uses, then
+  `GET /api/calendar`, and asserts the Riverbend ticket is due
+  2026-01-09. Without this, a seed helper that wrote the wrong date
+  would only fail in a live compose stack.
+
+## [0.21.0] - 2026-08-13
+
+### Added
+
+- `make seed` now inserts the synthetic Riverbend commitment post
+  (`fixtures.ambiguous_commitment_post`, created_at 2026-01-05) plus
+  one open dated ticket (due 2026-01-09). Fresh stacks show a row on
+  the home-page Calendar instead of an empty panel that made 0.18.0's
+  surface look unfinished. Re-seed is idempotent. The empty-state copy
+  tells the buyer how to populate the calendar (Derive, or a ticket
+  with a due date) when there really are none.
+
+## [0.20.0] - 2026-08-13
+
+### Added
+
+- The post-detail Event Lineage panel now draws the same git-branch SVG
+  as the home page, scoped to that post's reconstruct group
+  (`subgraphForPost`). Opening an A-100 post shows the `rec-002` fork
+  and keeps B-200 out; the direct/indirect lists stay as the
+  why-linked distinction. Tests drive `subgraphForPost` on the
+  designed A-100 fixture.
+
+## [0.19.0] - 2026-08-13
+
+### Added
+
+- `fast-mlsirm` as a pinned git dependency (`pyproject.toml`) -- the
+  first of three staged slices toward weekly/monthly PU/team/project
+  reports (see [ADR 0003](docs/adr/0003-fast-mlsirm-report-integration.md)).
+  It already implements the brief's LLM-as-a-Judge -> IRT-row ->
+  Fixed-Item Parameter Calibration pipeline, provider-neutrally against
+  contextual-orchestrator; reused rather than reimplemented. No
+  product behavior change in this slice -- infra only.
+- `requires-python` moved from `>=3.10` to `>=3.12` to match
+  `fast-mlsirm`'s own floor. `backend/Dockerfile`'s build stage gained
+  a pinned, non-interactive `rustup` install (`build-essential` for
+  the linker), since `fast-mlsirm` ships a PyO3/maturin Rust core with
+  no fallback wheel. The pytest CI job installs the same toolchain
+  before `pip install -e ".[backend]"`. Verified the compiled
+  extension actually loads (not the NumPy parity fallback) both
+  locally and in a freshly built `backend` Docker image. TEPP stays
+  the temporal/event measurement platform -- this pin is the IRT
+  compute library, not a fork of TEPP.
+
+## [0.18.0] - 2026-08-13
+
+### Added
+
+- Customer commitment derivation and a calendar/to-do surface -- the
+  brief's "이슈 관리는 To Do 와 캘린더로 자동 등록" and "고객과의
+  약속에 관해서 LLM 자동 도출" items, unified into one design: a
+  derived commitment *is* the ticket that shows up on the calendar,
+  not a separate concept. `lineageweave/commitment_extraction.py`:
+  pluggable LLM client that decides whether a post contains a genuine
+  customer commitment (a promise with a deadline, not just any event)
+  and resolves relative phrases ("by next Friday") against a supplied
+  reference date. `POST /api/posts/{post_id}/derive-commitment`
+  (`post_admin`) runs it and, when found, persists an `issue_ticket`
+  with `due_date`/`commitment_summary` set; `has_commitment: false` is
+  a normal 200, not an error, since most posts have no commitment.
+  `GET /api/calendar` lists every dated ticket the account may see,
+  soonest first, ABAC-filtered per row like every other cross-post
+  endpoint.
+- `issue_ticket` gained `due_date` and `commitment_summary` columns
+  (folded into `migrations/0001_initial_schema.sql`, plus
+  `migrations/0003_ticket_commitment_calendar.sql` for upgrading an
+  existing volume, matching the `0002` precedent).
+- Frontend: a "Derive commitment" action in the ticket panel, due
+  dates shown on ticket rows, and a Calendar panel on the product home
+  page listing upcoming commitments across every visible post,
+  clickable through to the source post.
+- Verified with a real LLM call through a locally-launched
+  contextual-orchestrator instance: `ambiguous_commitment_post`'s
+  relative "by next Friday" deadline resolved to the correct absolute
+  date against a supplied reference date, and a second sentence that
+  merely *looks* date-adjacent (a past event) was correctly not
+  treated as a commitment.
+
+### Fixed
+
+- `due_date` was a `timestamptz`. Binding a Python `date` becomes
+  midnight in the session timezone, then `.date()` on the UTC
+  (or local) datetime shifts the calendar day -- "2026-01-09"
+  came back as "2026-01-08". The column is now a `date`, matching
+  the LLM's YYYY-MM-DD. A malformed string still surfaces as 422.
+- Derive used wall-clock `now()` as the reference date. Relative
+  phrases in a historical post ("by next Friday") must resolve against
+  the post's `created_at` (TimeML document creation time), or the
+  calendar entry lands on the Friday after the click, not the Friday
+  the commitment was made.
+- Re-deriving the same post stacked a second open ticket. The persist
+  path now upserts the existing open commitment ticket.
+- `GET /api/calendar` included closed tickets, so finished work still
+  looked upcoming. Closed rows are excluded; a dated ticket created
+  through the regular ticket API still appears while it is open.
+- The "Derive commitment" button was shown to accounts without
+  `post_admin` and then 403'd. It is now gated the same way as Extract
+  Keymen. Calendar rows also show the source post title.
+- Ticket SQL stopped interpolating a column-list constant via f-string
+  (`asyncpg-sqli` Semgrep). The column list is a static literal in
+  each query; values stay parameterized.
+
+## [0.17.0] - 2026-08-13
+
+### Added
+
+- Valkey as a real event queue -- not a cache, not a second database. It
+  had been running in `docker-compose.yml` since Phase 1 with nothing in
+  the codebase ever publishing or reading from it (the brief's "Event
+  Queue, not MQ" requirement was unfulfilled infrastructure).
+  `backend/app/activity_stream.py`: ticket create/status-change
+  mutations `XADD` an event onto a per-post stream
+  (`activity:{post_id}`, approximately trimmed to the most recent 1000
+  entries); `GET /api/posts/{post_id}/activity` reads it straight back
+  with `XREVRANGE` -- no consumer group, no background worker, the
+  smallest slice where Valkey is actually load-bearing.
+- Frontend: an Activity panel in the popup (list + manual Refresh),
+  genuinely wired to the new endpoint.
+- Verified through the real Docker Compose network end to end (not just
+  `pytest`): created and patched a ticket via the actual backend
+  container, confirmed the events on `GET .../activity`, and confirmed
+  directly with `valkey-cli XLEN` against the actual `valkey` container
+  that the stream is real.
+
+## [0.16.0] - 2026-08-13
+
+### Added
+
+- Affiliate tree and VOC evidence on the post-detail popup. The schema
+  already stored a self-referencing `corporate_entity` hierarchy and a
+  `voc_type_code`, but the buyer only saw a raw code and a comma-separated
+  affiliation list.
+  `lineageweave/affiliate_tree.py` builds the ancestor forest of the
+  organizations a post's Keymen actually touch (Bhattacharya & Getoor,
+  2007's candidate-generation stage: an unresolved name is its own root,
+  never a guessed parent). `lineageweave/voc_evidence.py` quotes the
+  sentence that names a classified organization (ACE mention extent;
+  Doddington et al., 2004) and returns nothing when the name is absent.
+- Backend: `GET /api/posts/{post_id}/affiliate-tree` and
+  `GET /api/posts/{post_id}/voc-evidence`, same RBAC+ABAC gate as every
+  other post endpoint. Seed now inserts Demo Group → Demo Corp so the
+  tree is visible on a freshly seeded stack.
+- Frontend: Affiliate tree panel, VOC excerpt block, clickable Keyman
+  (loads `GET /api/keymen/{id}/related`), Event Lineage links that open
+  the linked post, and Extract Keymen for `post_admin`.
+
+## [0.15.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 5: issue ticket management -- the one explicit
+  product-brief item with a schema table (`issue_ticket`) but no
+  implementation until now. `backend/app/issue_ticket_ingestion.py`:
+  CRUD helpers, deliberately not a pluggable-LLM channel -- ticket
+  status is a closed enum in `common_lookup_value`.
+  `GET`/`POST /api/posts/{post_id}/tickets`,
+  `PATCH /api/tickets/{issue_ticket_id}` -- same RBAC (`post_read` for
+  reads, `post_admin` for writes) + ABAC as every other post-scoped
+  endpoint. `PATCH` resolves the ticket's owning post first and
+  ABAC-checks that. An invalid `ticket_status_code` hits the real FK
+  and surfaces as 422.
+- Frontend: `IssueTicketPanel` in the popup -- list, create, and inline
+  status-update.
+
+### Fixed
+
+- `frontend/Dockerfile`: pid-file sed now matches `/run/nginx.pid`
+  (the alpine image's real path), so non-root `USER nginx` can start.
+
+## [0.14.0] - 2026-08-13
+
+### Added
+
+- Product home page renders the reconstruct DAG as a git-branch SVG
+  (same layout language as `web/index.html`), grouped by reconstruct's
+  `thread_group_key`. Branch points are orange, isolated roots stay
+  visible, and a node click opens that post. `GET /api/lineage` now
+  includes `group` from the same `reconstruct_group_key()` rebuild uses,
+  so the UI cannot split A-100 / B-200 differently than reconstruct.
+- `Rebuild lineage` on the home page for `post_admin` accounts
+  (`POST /api/lineage/rebuild`). Viewers still only see the graph.
+
+## [0.13.1] - 2026-08-13
+
+### Fixed
+
+- Rebuild no longer derives reconstruct `group_key`/`secondary_key` from
+  process unit or voc type. Those collapsed A-100/B-200 and
+  proj-alpha/proj-beta and lost the designed `rec-002` fork.
+  `source_post` now persists `thread_group_key` and
+  `secondary_grouping_key`; seed writes fixture `occurred_at` into
+  `created_at`. `test_rebuild_lineage_recovers_the_a100_fork` drives
+  rebuild on the same A-100+B-200 rows seed inserts.
+
+## [0.13.0] - 2026-08-13
+
+### Added
+
+- `POST /api/lineage/rebuild` (`post_admin`): runs `reconstruct()` over
+  every `source_post` and writes `post_lineage_edge`. Grouping keys are
+  persisted on the post (`thread_group_key` / `secondary_grouping_key`).
+- `GET /api/lineage`: ABAC-filtered `{nodes, edges}` in the same shape as
+  the stdlib demo server, so the product UI can show the reconstructed
+  DAG. The React home page now lists those parent→child edges.
+- `backend/tests/test_api.py::test_rebuild_lineage_recovers_the_a100_fork`
+  inserts the designed A-100 fixture as real `source_post` rows, rebuilds,
+  and asserts the fork is visible on both the graph and per-post lineage.
+
+## [0.12.0] - 2026-08-13
+
+### Added
+
+- `lineageweave/lineage_persistence.py`: flattens `reconstruct()` trees
+  into the `(parent, child, fused_score)` rows `post_lineage_edge`
+  stores. `tests/test_lineage_persistence.py` proves the designed A-100
+  fork (`rec-002` → `rec-003` and `rec-004`) is in that contract, and
+  that unrelated `rec-006` is not forced onto a parent.
+- `scripts/seed_demo_data.py` now inserts the synthetic A-100 / B-200
+  fixture posts and persists those reconstruct edges, so
+  `GET /api/posts/{id}/lineage` (the Event Lineage panel) is not empty
+  on a freshly seeded demo stack.
+
+## [0.11.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 4: the post-detail popup's remaining panels.
+  `lineageweave/post_summary.py` -- LLM-derived Korean summary, key
+  events (ACE-style, Doddington et al. 2004), and R&R (semantic role
+  labeling, Gildea & Jurafsky, 2002), via contextual-orchestrator.
+  `lineageweave/post_chat.py` -- in-popup chat as retrieval-augmented
+  generation (Lewis et al., 2020): an explicit retrieve step
+  (`backend/app/post_chat_ingestion.py` assembles a post's own content
+  plus its Event-Lineage-linked posts, direct and Knowledge-Graph-
+  indirect, as numbered sources) then a reason-and-cite step (the model
+  answers using only those sources and reports which it drew from).
+- Backend: `GET /api/posts/{post_id}/lineage` (direct vs. indirect links,
+  kept as two separate lists), `GET /api/posts/{post_id}/summary`,
+  `POST /api/posts/{post_id}/chat` -- all RBAC+ABAC-gated the same way as
+  every other post endpoint, verified end to end against a live
+  Postgres + Keycloak + contextual-orchestrator stack (including through
+  the actual Docker-built images, not just the FastAPI TestClient).
+  Caught and fixed a real bug while building the chat's retrieve step:
+  `backend/app/knowledge_graph.py::load_visible_subgraph` only loads
+  edges among an *already-known* post set -- it does not itself discover
+  sibling posts sharing a mentioned person (its only prior caller,
+  `related_for_person`, pre-resolves that full set itself before calling
+  it). `find_linked_post_ids` now does that expansion first; regression-
+  tested (`test_post_chat_cites_a_post_linked_only_via_a_shared_keyman`).
+- Frontend: `frontend/src/App.tsx`'s popup gained real Summary, Event
+  Lineage (direct/indirect visually distinguished), Keyman, Counterparty,
+  and Chat sections, plus a real sliding evidence panel (`EvidencePanel`,
+  CSS animation, not a mock) that opens the cited source post's actual
+  content when a citation chip is clicked. Fixed a real, pre-existing
+  test-infrastructure bug found while adding more tests to the same file:
+  `@testing-library/react`'s automatic per-test cleanup never actually
+  ran (this project's `vite.config.ts` deliberately doesn't set
+  `test.globals`, which auto-cleanup silently depends on), so DOM from
+  one test was bleeding into the next; `src/setupTests.ts` now registers
+  `cleanup` in an explicit `afterEach`.
+- `docs/adr/0002-figma-access-boundary.md`: the referenced Figma frame is
+  genuinely reachable in this environment, but its own cover page is the
+  source organization's real confidential content (and it does not yet
+  contain a frame for this screen) -- the popup is built from the
+  product brief's text instead, named as an explicit, honest gap rather
+  than either faking a "Figma-matched" claim or stalling on it.
+- New citations: See, Liu, & Manning (2017); Gildea & Jurafsky (2002);
+  Lewis et al. (2020).
+
+## [0.10.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 3: `lineageweave/entity_relationship_classification.py`
+  -- LLM classification (via contextual-orchestrator, `mode="route"`) of a
+  named organization's relationship to the post author's org into the
+  product's own six-way vocabulary (`rel_voc`/`rel_vom`/`rel_vop`/
+  `rel_vocc`/`rel_voco`/`rel_vos` -- `rel_`-prefixed because
+  `common_lookup_value.lookup_code` is unique globally across categories
+  and bare `voc`/`vom` were already claimed by `source_post.voc_type_code`'s own
+  category). Proven with a real LLM call against a genuinely hard fixture
+  (`fixtures.ambiguous_entity_relationship_post`): an organization that is
+  both a repeat customer and a newly-competing division in the same post.
+- `lineageweave/corporate_hierarchy_resolution.py` -- similarity-based
+  resolution of a free-text organization name to an existing
+  `corporate_entity` row (Bhattacharya & Getoor, 2007's candidate-
+  generation/blocking stage of collective entity resolution, honestly
+  documented as that stage and not the full joint-inference version).
+  Wired into `backend/app/keyman_ingestion.py`, replacing the exact-
+  case-insensitive-match lookup it had before. `tests/test_corporate_hierarchy_resolution.py`
+  proves it against the same "Acme Group / Acme Electronics Korea / Acme
+  Electronics Gwangju Plant" hierarchy `tests/test_schema.py` already
+  uses: an abbreviation and a trailing legal suffix still resolve to the
+  right entity (not a sibling with a similar name), and a genuinely
+  unrelated organization resolves to `None`, not a guess.
+- `tests/test_indirect_lineage_linking.py`: demonstrates the Knowledge
+  Graph layer finds a real relation `lineageweave.reconstruct` has no
+  mechanism to find at all -- two posts in different `reconstruct.py`
+  groups (structurally never compared against each other) still surface
+  as related once they share a Keyman, via `random_walk_with_restart`.
+- Backend: `GET /api/posts/{post_id}/counterparties`, and
+  `POST /api/posts/{post_id}/extract-keymen` now also classifies and
+  persists each extracted Keyman's affiliated organizations'
+  relationships (`backend/app/entity_relationship_ingestion.py`), all on
+  the same RBAC+ABAC gate as the post/Keyman endpoints. Proven end to end
+  against a live Postgres + Keycloak + contextual-orchestrator stack.
+- `docs/lineage-bi-research-notes.md`: added Zelenko, Aone, & Richardella
+  (2003) (relation extraction); moved Bhattacharya & Getoor (2007) from
+  "staged for later phases" into a real section now that it grounds
+  shipped code.
+
+## [0.9.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 2 continues: `lineageweave/keyman_extraction.py`
+  extracts two-sided Keymen (our-side vs counterparty, 0..N organization
+  affiliations) from a post's title+body. The live client calls
+  contextual-orchestrator with `mode="route"` -- never a raw LLM API.
+  `NullKeymanExtractionClient` stays unavailable rather than inventing
+  mentions.
+- `lineageweave.knowledge_graph.knowledge_graph_edges_for_post` populates
+  the three Phase 2 edge kinds (person<->post mention, person<->corporate
+  entity affiliation, person<->person co-mention) as typed
+  `knowledge_graph_edge` specs. RWR then runs on that graph.
+- Backend endpoints `GET /api/posts/{post_id}/keymen`,
+  `GET /api/keymen/{person_id}/related`, and
+  `POST /api/posts/{post_id}/extract-keymen`, RBAC+ABAC-gated the same
+  way as post detail: a Keyman who is only mentioned on another corp's
+  private post 403s, related-node traversal never returns those hidden
+  posts, and extraction is `post_admin` (a write with a real LLM cost).
+  Persist goes through `backend/app/keyman_ingestion.py` into
+  `cataloged_person` / `person_affiliation` / `post_person_mention` /
+  `knowledge_graph_edge`. Directed RWR sinks teleport remaining walk
+  mass back to the start node so relevance stays a distribution.
+- `fixtures.ambiguous_keyman_post`: a synthetic, non-templated workshop
+  follow-up used by the parser tests and the real-orchestrator Keyman test.
+
+## [0.8.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 2 begins (Phase 1 complete): `lineageweave/knowledge_graph.py`
+  -- random walk with restart (Tong, Faloutsos, & Pan, 2006) giving every
+  node in a Knowledge Graph a continuous relevance score from a starting
+  node, plus `select_related_nodes`'s per-node adaptive relevance-ratio
+  cutoff. No hop-count constant appears anywhere in the algorithm -- the
+  same ratio threshold naturally yields a larger related-set for a
+  well-connected node than a sparse one, which is the actual product
+  requirement ("각 Node와 Node마다 Depth가 다를 수 있다").
+- `tests/test_knowledge_graph.py`: proves this against a synthetic graph
+  built so the correct answer is known by construction -- a disconnected
+  node scores exactly 0, symmetric spokes score identically, and a
+  well-connected "hub" node's adaptive related-set (5 nodes) is
+  measurably larger than a sparse "loner" node's (1 node) under the same
+  threshold.
+- `docs/lineage-bi-research-notes.md`: Tong et al. (2006) moved from
+  "staged for later phases" into a real "Knowledge Graph traversal"
+  section now that it grounds shipped code.
+
+## [0.7.0] - 2026-08-13
+
+### Added
+
+- `frontend/`: React + Vite + TypeScript, pinned Node via `mise.toml`,
+  pnpm via Corepack -- a real client, not mocked and not static HTML.
+  `react-oidc-context` drives an actual Authorization Code redirect
+  through Keycloak; the post list and detail popup call the FastAPI
+  backend over real `fetch()` with the token Keycloak issued.
+  `src/App.test.tsx` covers the login-redirect and fetch-then-render
+  paths (`useAuth` mocked -- the real OIDC round-trip is proven
+  elsewhere, by `scripts/smoke_test_oidc.py` and `backend/tests/test_api.py`).
+- `frontend/Dockerfile` + `nginx.conf`: two-stage build (`pnpm run build`
+  then nginx serving the static bundle) added as docker-compose's fourth
+  service, `VITE_*` config baked in at build time from the same `.env`
+  ports every other service uses. Both stages pin the base image by
+  digest; the runtime stage declares `USER nginx` and listens on 8080
+  so the master process does not need root to bind a port.
+- `backend/app/main.py`: `CORSMiddleware`, scoped to exactly the
+  frontend's origin(s) (`FRONTEND_ORIGINS`), `GET` only, `Authorization`
+  header only -- verified with a real cross-origin preflight + GET against
+  the live stack, not just unit-tested in isolation.
+- `.github/workflows/tests.yml`: added a `frontend` job (lint, test,
+  build) alongside the existing Python `pytest` job.
+
+### Fixed
+
+- Keycloak's `lineageweave-frontend` client (`docker/keycloak/realm-export.json`)
+  now allows both the Vite dev-server origin (`:5173`) and the
+  docker-compose-served frontend's origin (`:15173`) as redirect URIs and
+  web origins -- the login redirect only worked from one of the two
+  before this.
+
+## [0.6.0] - 2026-08-13
+
+### Added
+
+- `backend/`: a FastAPI app connecting directly to PostgreSQL (`asyncpg`,
+  no ORM, no file-backed DB). OIDC bearer-token login verified against a
+  live Keycloak JWKS (`backend/app/auth.py`, fetched via
+  `lineageweave.http_client` rather than PyJWKClient/`urllib`); RBAC
+  (`post_read` permission via role membership) plus row-level ABAC
+  (private `source_post` rows scoped to the requesting account's
+  affiliated corporate entity) enforced on `GET /api/posts` and
+  `GET /api/posts/{post_id}` (`backend/app/main.py`).
+- `backend/tests/test_api.py`: real-integration tests -- a genuine access
+  token from a live Keycloak, verified against a throwaway migrated
+  Postgres database. Proves both the allow and the deny path: a private
+  post scoped to a *different* corporate entity is excluded from the list
+  and 403s on direct fetch; a forged token is rejected; a missing token is
+  401. Skipped unless both a local PostgreSQL and Keycloak are reachable.
+- `scripts/seed_demo_data.py` (`make seed`): seeds synthetic corp/PU/
+  account/`source_post` rows keyed to the *real* subject ids Keycloak's
+  admin REST API reports for the two demo users -- not a locally-fabricated
+  guess at what those ids might be. Talks to Keycloak through
+  `lineageweave.http_client` (`post_form` / `get_json_list`), never
+  `urllib.request.urlopen`.
+- `backend/Dockerfile`: `python:3.12-slim` pinned by digest, runtime
+  `USER appuser` (DS-0002).
+- `migrations/0001_initial_schema.sql`: added `corporate_entity.
+  corporate_entity_code` (unique short code, e.g. `DEMO-CORP-01`) -- the
+  column the login-time `corp_code` claim actually maps to; the original
+  Phase 1 migration (still unmerged) only had the human-readable
+  `entity_name`. Postgres's app database is now auto-migrated with this
+  exact file on first `docker compose up` (`docker/postgres-init/Dockerfile`),
+  so what's tested and what ships never drift apart.
+- docker-compose.yml's default host ports moved off Postgres/Redis/common
+  local-dev ports entirely (15432, 16379, 18080, 18420) -- found during
+  this work that a colliding already-running service on a container's
+  published port can silently answer curl/psql requests instead of the
+  container, with no error; picking non-default ports avoids that
+  ambiguity outright rather than relying on operators noticing.
+
+### Fixed
+
+- JWKS fetch and the demo seeder go through `lineageweave.http_client`
+  instead of `PyJWKClient` / `urllib.request.urlopen`.
+
+## [0.5.0] - 2026-08-13
+
+### Added
+
+- `docker-compose.yml`: PostgreSQL, Valkey, and a real Keycloak OIDC
+  provider, genuinely functional (not a stub/mocked adapter). `make up`
+  brings up all three from a clean checkout; `make smoke` runs
+  `scripts/smoke_test_oidc.py`, which logs in as a synthetic demo user
+  seeded by `docker/keycloak/realm-export.json`, fetches Keycloak's live
+  JWKS, and cryptographically verifies the returned JWT's RS256 signature,
+  issuer, and `corp_code`/`pu_code` custom claims -- a real round-trip
+  proof, not a "the container started" check.
+- `docker/postgres-init/` and `docker/keycloak/`: both services are `build:`
+  targets (Dockerfiles that `COPY` in the keycloak-db init script and the
+  realm seed) rather than bind mounts, so the images are self-contained and
+  reproducible on any Docker host or CI runner.
+- Keycloak stores its own state in a second database (`keycloak`) on the
+  same PostgreSQL instance -- one running database service for the whole
+  stack, no second file-backed store.
+- `.env.example` documents the (already-defaulted) compose variables,
+  including how to remap host ports if 5432/6379/8080 are already taken
+  locally.
+
+### Fixed
+
+- Dockerfiles declare an explicit non-root `USER` (DS-0002) and pin
+  `postgres:16-alpine`, `quay.io/keycloak/keycloak:26.0`, and
+  `valkey/valkey:8-alpine` by digest.
+- `scripts/smoke_test_oidc.py` talks to Keycloak through
+  `lineageweave.http_client` (`get_json` / `post_form`) instead of
+  `urllib.request.urlopen`, so the same `file://` allowlist used by the
+  library clients applies to the OIDC smoke path.
+
+## [0.4.0] - 2026-08-13
+
+### Added
+
+- Milestone 4, Phase 1 begins: LineageWeave's product schema.
+  `migrations/0001_initial_schema.sql` -- a 3NF PostgreSQL schema
+  (snake_case, 2+ word object names) covering accounts (corp/PU code as
+  attributes, not the login key), a shared `common_lookup_value` ENUM
+  table, posts + visibility, ABAC/RBAC, VOC-type and entity-relationship
+  classification, Keyman (`cataloged_person` + N:N `person_affiliation`), a
+  `knowledge_graph_edge` table, `issue_ticket`, and a self-referencing
+  `corporate_hierarchy` via `corporate_entity.parent_entity_id`.
+- `docs/adr/0001-demo-identity-and-data-boundary.md`: the identity/data
+  scope decision for this expansion -- real infrastructure (Postgres,
+  Valkey, a real OIDC provider), synthetic identities and content, because
+  Keyman extraction catalogs real named individuals (including
+  non-consenting external counterparties) and a real production identity
+  provider would re-identify the source organization through account/data
+  structure even with zero literal company-name strings in source files.
+- `tests/test_schema.py`: real-database tests (skipped without a
+  reachable PostgreSQL server) proving the migration applies cleanly, a
+  multi-level corporate-hierarchy recursive query returns the right
+  shape, and an invalid lookup code is genuinely rejected by a foreign
+  key -- caught and fixed a real bug in the process (an accidental
+  `deferrable initially deferred` on one FK silently weakened its
+  integrity check within a transaction).
+- New citations staged for Phase 2/3: Tong et al. (2006, random walk with
+  restart -- Knowledge Graph per-node traversal depth) and Bhattacharya &
+  Getoor (2007, collective entity resolution -- corporate hierarchy).
+
+## [0.3.0] - 2026-08-13
+
+### Fixed
+
+- Embedding, adjudication, and vision clients now POST through a shared
+  `http_client.post_json` helper that allowlists `http`/`https` and never
+  calls `urllib.request.urlopen`. That closes the `file://` read concern
+  Semgrep's `dynamic-urllib-use-detected` rule was flagging on the
+  operator-configured base URLs. HTTPS posts wrap the connected socket
+  with a certifi-backed `SSLContext` instead of constructing
+  `http.client.HTTPSConnection`, so certificate verification is explicit
+  on the Python 3.10+ runtime this project requires.
+
+### Added
+
+- `lineageweave/chunking.py`: semantic-unit chunking so the embedding
+  channel compares meaning-identifiable units instead of whole flattened
+  documents -- `chunk_by_paragraph` (Hearst, 1997, TextTiling subtopic
+  boundaries), `chunk_by_sentence`, `chunk_by_dom` (WHATWG HTML Living
+  Standard sectioning/flow block elements), and `chunk_by_conversation_turn`
+  (RFC 5322 sender/receiver boundaries).
+- `embedding_client.chunked_max_similarity`: chunks two documents, embeds
+  every chunk, and returns the single highest-scoring pair -- the standard
+  passage-retrieval strategy for "a relevant unit is buried in a longer
+  document." Degrades to plain whole-text embedding for any document that
+  chunks to zero or one piece (this project's real short-title dataset
+  behaves exactly as it did before chunking existed).
+- Real-provider test proving chunking works, not just that it type-checks:
+  a short relevant paragraph buried inside a longer synthetic document
+  scored higher via `chunked_max_similarity` than via whole-document
+  embedding, against the live embedding provider.
+- `docs/lineage-bi-research-notes.md`: new "Chunking" section with the
+  four units' grounding and an explicit, honest note that this project's
+  unseen dataset's only free-text field is too short to need chunking in
+  practice -- the module exists for richer content sources (e.g. the raw
+  MHTML artifacts that dataset's records were derived from).
+- `lineageweave/image_content.py`: pluggable vision channel for base64
+  images embedded in DOM content -- real OCR (Li et al., 2023, TrOCR) and
+  object recognition/tagging (Radford et al., 2021, CLIP) via
+  `OpenAiCompatibleVisionClient`, same never-fake-a-missing-channel
+  discipline as the embedding/adjudication clients. `chunk_by_dom` now
+  extracts embedded images as `"image"` chunks interleaved with text
+  chunks in true document order, so an image's position relative to its
+  surrounding text is preserved and reconstructable.
+- Real-provider test proving OCR works, not just that it type-checks: a
+  real PNG generated with real rendered text (not a fixture file) was
+  read back correctly by `OpenAiCompatibleVisionClient` against the live
+  vision-capable model.
+- `docs/image-content-schema.md`: proposed DB schema (snake_case, 2+ word
+  object names) for persisting and searching extracted image content,
+  designed so a text/tag search hit stays traceable to which document and
+  which position produced it, and so the same image (by content hash) is
+  never *stored* twice (the primary key guarantees that part). Avoiding a
+  duplicate vision-provider *call* for two concurrent ingests of the same
+  new image is a separate concern the schema documents but does not solve
+  by itself -- a real write path still needs an atomic claim/lease step.
+
+## [0.2.0] - 2026-08-13
+
+### Added
+
+- ADR-0016 grounding: `docs/lineage-bi-research-notes.md` now cites
+  Doddington et al. (2004, ACE) and Anagnostopoulos et al. (2013, CHRONOS)
+  alongside Allan (2002, TDT), and explains how `reconstruct.py` maps onto
+  TEPP's three-layer event-intelligence separation (mention/instance,
+  calibrated detection, temporal-consistency).
+- `tests/test_real_provider_integration.py`: opt-in (env-var-gated, skipped
+  by default so a credential-free clone stays green) tests proving
+  `OpenAiCompatibleEmbeddingClient` and `ContextualOrchestratorAdjudicationClient`
+  work end-to-end against real providers, not just that their interfaces
+  are satisfiable by a stub.
+
+### Fixed
+
+- `embedding_client.py` / `adjudication_client.py`: both real-provider HTTP
+  clients now build their SSL context from `certifi`'s CA bundle instead of
+  the platform default. Some interpreter distributions (observed: a
+  standalone uv-managed CPython on macOS) don't reliably inherit the OS
+  trust store, which made every real-provider call fail closed with
+  `CERTIFICATE_VERIFY_FAILED` even against a validly, publicly-trusted
+  certificate. Full chain validation still applies -- nothing is weakened,
+  the bundle source just changed.
+
+## [0.1.0] - 2026-08-13
+
+### Added
+
+- Initial prototype: `reconstruct()` pipeline (group → bounded candidate
+  window → multi-channel fusion via RankWeave → tree assembly via
+  ThreadWeave).
+- Four scoring channels: `temporal`, `secondary_key`, `text`
+  (dependency-free stand-in for embedding-cosine similarity), and an
+  optional `llm` channel via a pluggable `AdjudicationClient`.
+- `ContextualOrchestratorAdjudicationClient`, calling
+  [contextual-orchestrator](https://github.com/ContextualWisdomLab/contextual-orchestrator)'s
+  `mode="verify"` with `reasoning_effort="high"`.
+- `OpenAiCompatibleEmbeddingClient` for a real embedding-cosine text
+  channel against any OpenAI-compatible `/v1/embeddings` endpoint.
+- `TeppClient` / `AnalysisRunRequest`: validated wire shape matching
+  [TEPP](https://github.com/ContextualWisdomLab/TEPP)'s published
+  `analysis_run_request_v1.json` schema, pluggable transport (fails closed
+  with `TeppNotAvailable` until TEPP ships a live HTTP endpoint).
+- Minimum fused-score floor so weakly-related records surface as their own
+  root instead of being force-attached to the best of a bad set of
+  candidates.
+- Stdlib HTTP demo server (`lineageweave/server.py`) and a self-contained
+  SVG DAG viewer (`web/index.html`), no build step or external script
+  dependency.
+- Synthetic-only demo dataset (`lineageweave/fixtures.py`).
+- `docs/lineage-bi-research-notes.md`: the literature this design is
+  grounded in, APA 7th.

From 5410f76d22dc890954ec87a9eef85319aad8c3a0 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:06:54 -0700
Subject: [PATCH 09/11] test(ask): cover modal exit focus and source transition

---
 .../components/AskEvidenceLayerPopup.test.tsx | 23 +++++++++++++++++--
 1 file changed, 21 insertions(+), 2 deletions(-)

diff --git a/frontend/src/components/AskEvidenceLayerPopup.test.tsx b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
index dac6a5147..7bed0b01b 100644
--- a/frontend/src/components/AskEvidenceLayerPopup.test.tsx
+++ b/frontend/src/components/AskEvidenceLayerPopup.test.tsx
@@ -77,19 +77,22 @@ describe("AskEvidenceLayerPopup", () => {
     expect(onClose).toHaveBeenCalledTimes(3);
   });
 
-  it("opens the cited post from the layer", async () => {
+  it("closes the evidence layer before opening the cited post", async () => {
+    const onClose = vi.fn();
     const onOpenPost = vi.fn();
     render(
       <AskEvidenceLayerPopup
         {...baseProps}
         facts={[]}
         images={[]}
-        onClose={vi.fn()}
+        onClose={onClose}
         onOpenPost={onOpenPost}
       />,
     );
     await userEvent.click(screen.getByRole("button", { name: "Open post: Checkout error follow-up" }));
+    expect(onClose).toHaveBeenCalledTimes(1);
     expect(onOpenPost).toHaveBeenCalledWith("post-demo-public");
+    expect(onClose.mock.invocationCallOrder[0]).toBeLessThan(onOpenPost.mock.invocationCallOrder[0]);
   });
 
   it("moves initial focus onto the dialog panel", () => {
@@ -114,4 +117,20 @@ describe("AskEvidenceLayerPopup", () => {
     await userEvent.tab({ shift: true });
     expect(openPostButton).toHaveFocus();
   });
+
+  it("returns focus to the element that invoked the modal when the layer unmounts", () => {
+    const opener = document.createElement("button");
+    opener.textContent = "View evidence";
+    document.body.append(opener);
+    opener.focus();
+
+    const { unmount } = render(
+      <AskEvidenceLayerPopup {...baseProps} facts={[]} images={[]} onClose={vi.fn()} onOpenPost={vi.fn()} />,
+    );
+    expect(screen.getByRole("dialog")).toHaveFocus();
+
+    unmount();
+    expect(opener).toHaveFocus();
+    opener.remove();
+  });
 });

From 17c2f085d74c5ac22874878499d9051bf44f6752 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sat, 22 Aug 2026 03:07:19 -0700
Subject: [PATCH 10/11] fix(ask): restore focus when evidence modal exits

---
 .../src/components/AskEvidenceLayerPopup.tsx   | 18 +++++++++++++++++-
 1 file changed, 17 insertions(+), 1 deletion(-)

diff --git a/frontend/src/components/AskEvidenceLayerPopup.tsx b/frontend/src/components/AskEvidenceLayerPopup.tsx
index fe3c68c43..4aff6e88e 100644
--- a/frontend/src/components/AskEvidenceLayerPopup.tsx
+++ b/frontend/src/components/AskEvidenceLayerPopup.tsx
@@ -51,9 +51,16 @@ export function AskEvidenceLayerPopup({
   const factsHeadingId = useId();
   const imagesHeadingId = useId();
   const panelRef = useRef<HTMLDivElement>(null);
+  const restoreFocusOnUnmountRef = useRef(true);
 
   useEffect(() => {
+    const previouslyFocused = document.activeElement instanceof HTMLElement ? document.activeElement : null;
     panelRef.current?.focus();
+    return () => {
+      if (restoreFocusOnUnmountRef.current && previouslyFocused?.isConnected) {
+        previouslyFocused.focus();
+      }
+    };
   }, []);
 
   useEffect(() => {
@@ -96,6 +103,15 @@ export function AskEvidenceLayerPopup({
     return () => document.removeEventListener("keydown", handleKeyDown);
   }, [onClose]);
 
+  function handleOpenPost() {
+    // Opening the full post is a workflow transition rather than returning to
+    // the invoking citation, so do not restore focus to the old trigger while
+    // the destination surface is being mounted.
+    restoreFocusOnUnmountRef.current = false;
+    onClose();
+    onOpenPost(postId);
+  }
+
   return (
     <div className="popup-backdrop" onClick={onClose}>
       <div
@@ -138,7 +154,7 @@ export function AskEvidenceLayerPopup({
             </ul>
           </section>
         ) : null}
-        <button type="button" className="keyman-select" onClick={() => onOpenPost(postId)}>
+        <button type="button" className="keyman-select" onClick={handleOpenPost}>
           {tf("Open post: {label}", { label: postTitle })}
         </button>
       </div>

From eb77477948aca956b69683b1f961bd22563fb5b6 Mon Sep 17 00:00:00 2001
From: Seongho Bae <me@seonghobae.me>
Date: Sun, 23 Aug 2026 11:06:14 +0900
Subject: [PATCH 11/11] fix(frontend): drop stray pre-login AdminPanel, restore
 return-URL persistence

The pre-login screen rendered AdminPanel with a possibly-undefined
accessToken (a TS6192/TS2322 build break) and never used the persisted
return-URL helpers on the login redirect. Same fix as LineageWeave#456
on main, applied here since this stack predates that fix.
---
 frontend/src/App.tsx | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/frontend/src/App.tsx b/frontend/src/App.tsx
index ccfcaaed4..8fff1651b 100644
--- a/frontend/src/App.tsx
+++ b/frontend/src/App.tsx
@@ -4638,7 +4638,8 @@ export default function App({ showLabPanels = false }: { showLabPanels?: boolean
             </div>
             <div className="login-controls">
               <button className="btn-primary" onClick={() => {
-                const returnUrl = window.location.pathname + window.location.search;
+                const returnUrl = returnUrlFromLocation();
+                rememberOidcReturnUrl(returnUrl);
                 void auth.signinRedirect({ state: { returnUrl } });
               }}>
                 {t("Log in")}
@@ -4648,7 +4649,6 @@ export default function App({ showLabPanels = false }: { showLabPanels?: boolean
               <small>Enterprise SSO Authentication</small>
             </div>
           </div>
-          {destination === "admin" ? <AdminPanel currentBrandName={brandName} onBrandNameChange={setBrandName} accessToken={accessToken} /> : null}
       </main>
         <footer className="app-footer" role="contentinfo">
           <div className="app-footer-title">