feat: wire Claude Workflow gap readiness + ship real D1 receipts demo

.agent/demos/d1-receipts/NON-CLAIMS.md

@@ -0,0 +1,5 @@
+# Non-claims
+
+- This packet does not prove every sentence in PR #3282's body is independently verified -- only the four claims explicitly checked in report.md are scored; claim 4 is explicitly scored `not_supported`.
+- This packet does not establish that `session_refs` correctly resolves every PR reference archive-wide -- only that it resolves this one case with structural evidence.
+- This packet does not reproduce on the public seed corpus (seed 1843); it requires read-only access to the live archive and the `Sinity/polylogue` GitHub history.

.agent/demos/d1-receipts/PROMPT.md

@@ -0,0 +1,67 @@
+# D1 "The Receipts": Claim-vs-Evidence on a Real Merged PR
+
+Predeclaration receipt: `artifact:d1-receipts-predeclaration`.
+
+Pick a real merged, agent-authored PR from this repository. Resolve it to
+its authoring/dispatch session **structurally** — via `session_refs`
+(kind=`pull_request`), not by regex-scanning message prose or a time-window
+heuristic. Then check specific sentences from the PR body against that
+session's own recorded tool_use/tool_result blocks: does the evidence
+actually support the claim, or is the claim resting on the PR body's own
+prose with nothing underneath it?
+
+Product primitives only: `session_refs` (the typed evidence table wired by
+PR #3425/#3431), `polylogue read --view correlation`, and structural SQL
+reads over `blocks`/`session_refs` for citation (mirroring the exact
+read-only query style PR #3392 and PR #3282 themselves used in their own
+Verification sections — this demo does not invent a new access pattern).
+
+## Steps
+
+1. Resolve PR → session structurally:
+   ```sql
+   SELECT session_id, repo, ref_number, url
+   FROM session_refs
+   WHERE kind = 'pull_request' AND repo = 'Sinity/polylogue' AND ref_number = 3282;
+   ```
+   Cross-check the same resolution through the CLI's own read surface:
+   `polylogue find "id:<session_id>" then read --view correlation --format json`
+   (this is the surface PR #3425/#3431 wired `session_refs` into —
+   `insights/session_commit.py:build_correlation_result` and
+   `insights/correlation_view.py`).
+
+2. Fetch the PR body from GitHub (`gh pr view 3282 --json body`) and pull
+   out individually falsifiable sentences — not the whole prose block, each
+   claim on its own.
+
+3. For each claim, search the resolved session's own `blocks` rows
+   (`tool_use`/`tool_result`, joined by `tool_id`) for structural evidence:
+   an exact command, an exact exit code, an exact pytest summary line. A
+   claim with no matching block is marked **not independently verified in
+   this session** — never silently upgraded to "supported" because the PR
+   body asserts it.
+
+4. Render the two columns: claimed sentence | observed block evidence
+   (drillable via the cited `block:` ref), with an explicit status per row.
+
+## Note on this run
+
+This demo's session turned out to be a **merge-conductor** session: its own
+`blocks` are almost entirely `Bash` (53 of 56 tool_use blocks) plus 3 `Read`
+calls — zero `Edit`/`Write` tool_use. The actual file edits for PR #3282
+happened in separately dispatched worker sessions across multiple git
+worktrees (`/realm/worktrees/polylogue-membership-head*`); this session
+orchestrates `git`, `gh pr create`, and `devtools test`/`devtools verify`
+invocations across those worktrees and stitches the result into one PR.
+
+This is itself a real, useful finding, not a inconvenience to hide: the PR
+body's own Verification section names a 7-file `devtools test` invocation
+("`devtools test tests/unit/sources/test_live_batch_support.py ...` — all
+passing, **see individual commit messages for per-commit pass counts**") —
+its own parenthetical admits the aggregate command was never run as one
+shot. Searching this session's blocks confirms it: the 7-file string only
+appears inside the `gh pr create --body` tool_input (i.e. inside the PR body
+text itself), never as an actual invoked command. That specific claim is
+marked **not independently verified in this session** in `report.md` and
+`checks.json` — precisely the honesty discipline this packet exists to
+enforce, applied to itself.

.agent/demos/d1-receipts/evidence.ndjson

@@ -0,0 +1,11 @@
+{"ref": "artifact:d1-receipts-evidence", "cited_for": "Demo Packet v2 receipt root", "verified_via": "committed evidence.ndjson"}
+{"ref": "session:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc", "cited_for": "session_refs kind=pull_request resolves this session to Sinity/polylogue#3282", "verified_via": "sqlite3 index.db: SELECT session_id,repo,ref_number,url FROM session_refs WHERE kind='pull_request' AND repo='Sinity/polylogue' AND ref_number=3282"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:ae9a4788-4bf6-4b89-b01e-90f10e622981:0", "cited_for": "the gh pr create tool_use --body text is byte-identical to the PR body later fetched live via gh pr view 3282, including the exact bullet claims checked below", "verified_via": "diff of tool_input command against `gh pr view 3282 --repo Sinity/polylogue --json body`"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:a98d94d0-831f-4cf6-946b-40bf99179283:0", "cited_for": "gh pr create tool_result: https://github.com/Sinity/polylogue/pull/3282 -- confirms this session actually opened PR #3282, not merely referenced it", "verified_via": "tool_result text (single URL line)"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:bdacfb31-ecc7-4491-be2f-891f8bfb888b:0", "cited_for": "tool_use invoking `timeout 180 devtools verify --quick`, checking the PR claim 'devtools verify --quick -- pass'", "verified_via": "tool_input command text"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:9786b513-9ad0-4e25-ad3e-e18a0f60220f:0", "cited_for": "tool_result: structured verify-run JSON, every step exit=0, total_duration_s=32.99, exit_code=0 -- SUPPORTS the claim via structure, not a trusted pass/fail word in prose", "verified_via": "tool_result JSON body (per-step name/duration_s/exit array)"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:ebe785a2-de91-4de3-9b28-43bd1a8e9596:0", "cited_for": "tool_use invoking devtools test against tests/unit/maintenance/test_rebuild_index_bulk_build.py (+4 more files), checking the Solution-section claim 'rebuild_index bulk FTS materialization checkpoints progress'", "verified_via": "tool_input command text"}
+{"ref": "block:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc:59a1be2d-7881-4d64-997d-d055dd61aa74:0", "cited_for": "tool_result: pytest summary '123 passed in 8.05s', ok (12.2s) -- SUPPORTS the rebuild_index claim", "verified_via": "tool_result text (pytest summary line)"}
+{"ref": "polylogue-6mvg", "cited_for": "the PR body's own tracking-item reference ('Ref polylogue-6mvg'), cited verbatim inside the gh pr create --body text", "verified_via": "block:...ae9a4788... tool_input"}
+{"ref": "action.tool_use.count=53,Bash", "cited_for": "56 total tool_use blocks in this session: 53 Bash, 3 Read, 0 Edit, 0 Write -- this is a merge-conductor/orchestration session, not the direct file-editing session", "verified_via": "sqlite3 index.db: SELECT tool_name, count(*) FROM blocks WHERE session_id=... AND block_type='tool_use' GROUP BY tool_name"}
+{"ref": "action.negative-control.missing-7-file-devtools-test-invocation", "cited_for": "counterexample: the PR body's exact 7-file `devtools test tests/unit/sources/test_live_batch_support.py ...` string appears ONLY inside the gh-pr-create --body text (i.e. inside the PR body itself), never as an actually-invoked command in this session's blocks -- 0 matching tool_use rows when the gh-pr-create block is excluded", "verified_via": "sqlite3 index.db: SELECT count(*) FROM blocks WHERE session_id=... AND block_type='tool_use' AND tool_input LIKE '%test_live_batch_support.py%' AND tool_input NOT LIKE '%gh pr create%' -- returns 0"}

.agent/demos/d1-receipts/finding.yaml

@@ -0,0 +1,6 @@
+archive_cursor: "live-archive:/realm/db/polylogue (read-only, file:...?mode=ro)"
+measure_version: demo-packet-v2
+commit_sha: 59744a30bf461a587cc679ee62a432e8cd2cf82a
+sample_frame_predicate: "session_refs WHERE kind='pull_request' AND repo='Sinity/polylogue' AND ref_number=3282, resolving to claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc"
+run_date: "2026-07-31"
+claim: "session_refs typed pull_request evidence resolves a real merged PR to its authoring/dispatch session, and specific PR-body verification sentences can be checked against that session's own recorded blocks -- with unsupported claims marked as such, not silently believed"

.agent/demos/d1-receipts/queries.ndjson

@@ -0,0 +1,5 @@
+{"text": "SELECT session_id, repo, ref_number, url FROM session_refs WHERE kind='pull_request' AND repo='Sinity/polylogue' AND ref_number=3282", "lowered_spec": {"unit": "session_ref", "predicate_kind": "structural_equality", "table": "session_refs"}}
+{"text": "find \"id:claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc\" then read --view correlation --format json", "lowered_spec": {"unit": "session", "entry": "id", "view": "correlation"}}
+{"text": "SELECT tool_name, count(*) FROM blocks WHERE session_id='claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc' AND block_type='tool_use' GROUP BY tool_name ORDER BY 2 DESC", "lowered_spec": {"unit": "block", "pipeline_stages": ["group:tool_name", "count"]}}
+{"text": "SELECT tu.block_id, tr.text FROM blocks tu JOIN blocks tr ON tr.tool_id=tu.tool_id AND tr.block_type='tool_result' AND tr.session_id=tu.session_id WHERE tu.session_id='claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc' AND tu.tool_input LIKE '%devtools verify --quick%'", "lowered_spec": {"unit": "block", "predicate_kind": "join_tool_use_to_tool_result"}}
+{"text": "SELECT count(*) FROM blocks WHERE session_id='claude-code-session:5ecdb160-495a-4d9b-b80a-3a24886af8cc' AND block_type='tool_use' AND tool_input LIKE '%test_live_batch_support.py%' AND tool_input NOT LIKE '%gh pr create%'", "lowered_spec": {"unit": "block", "predicate_kind": "negative_control_count", "expected_result": 0}}