docs: recapture the knowledge screenshots

[larryro]

Summary

Regenerates the two knowledge screenshots through the real capture pipeline (bun run docs:screenshots) on the hermetic docs-demo stack, as the end-to-end proof that knowledge ingestion works again after the BM25 self-heal (#3212). Two commits:

• docs: recapture the knowledge screenshots — project-knowledge-files.webp, documents-list.webp, and the documents-list captions (en/de/fr × 3 pages). manifest.json is byte-identical (it stores route/viewport/dimensions only).
• chore(platform): seed the embedding model before the docs-shot uploads — the rig gap that had left every seeded document Failed (details below).

Why the documents were failed

The seeded files were not PANIC victims. All five .txt files and two of the three knowledge-entry documents had failed with "No embedding model is configured — set knowledge/embedding.json for this organization." — the demo org never had an embedding model, because the seeder never configured one (and its previous config dir lived inside a worktree that no longer exists). Only Onboarding kickoff window.md was the PANIC victim: knowledge-db row stuck at processing, file marked "Indexing was interrupted and did not finish. Re-upload the file to try again."

Recovery through the product path (no DB repair)

1. Settings → Data residency → Embedding model (org owner): provider e2e-mock, model text-embedding-3-small, width 1536 — exactly what the failure dialog tells an admin to do. The seeder now does this step through the same UI before the first upload (ensureEmbeddingModel).
2. Retry indexing per row — the project Knowledge tab's per-row button and the Documents badge's retry button. The seeder re-queues rows an earlier run left Failed the same way and waits for Indexed.
3. The interrupted row (Onboarding kickoff window.md) also recovered through Retry indexing — the ingest upserts ON CONFLICT (org_slug, file_id), so the stuck knowledge-db row was reset and re-chunked. Its rag_error text tells the user to re-upload; the retry button already suffices (product nit).

No dead-end: every failed document had a retry affordance. Nothing was touched in either database by hand.

Heal evidence (knowledge-db tale_knowledge, localhost:5433)

Probe	Before	After
select count(*) from private_knowledge.chunks	936 (all BGB.pdf, org test)	944 (+5 seeded .txt, +3 knowledge-entry .md; the deleted+re-uploaded brand-guidelines chunk nets zero)
pdb.verify_index('private_knowledge.idx_pk_chunks_bm25')	—	schema_valid t · index_readable t · checksums_valid t · segment_metadata_valid t (2 segments)
PANIC lines in newest log (postgresql-2026-09-04_071700.log)	0	0 (the only 2 ERROR lines are this morning's assertion left == right at 07:17/07:33 UTC, before the repair)
private_knowledge.documents (org northlight-labs)	1 row, processing	8 rows, all completed incl. Onboarding kickoff window.md at 09:26:24 UTC
app.file_metadata.rag_status (org)	8 × failed	8 × completed
mock gateway /v1/embeddings calls	—	9 (5 + 3 + 1 re-upload)

Boot-time verification of the same stack (the hermetic bun scripts/dev.ts swallows backend stdout, so this is an observation-only backend booted with the identical env on :3105):

[knowledge] the deployment-default knowledge database: BM25 index private_knowledge.idx_pk_chunks_bm25 is healthy (4 checks, 3.9 MB, 6 ms)
[knowledge] the deployment-default knowledge database: BM25 index public_web.idx_pw_chunks_bm25 is healthy (4 checks, 2.9 MB, 2 ms)

app.audit_logs holds two knowledge_index_repaired (success) events on private_knowledge.idx_pk_chunks_bm25 from 07:33 UTC today — the merged self-heal's live repair — and none since.

Rig changes (chore(platform))

• ensureEmbeddingModel seeds the org's embedding model through Settings → Data residency (mock provider, 1536-dim as the knowledge-db stores; EMBEDDING_DIMENSIONS is now exported from the mock so the width has one source). The mock-provider step moves ahead of the uploads to make the provider selectable.
• retryFailedIndexing + awaitIndexed: rows left Failed are re-queued through their own Retry indexing button; every uploaded/re-queued row gets a bounded, non-fatal indexing window (a stack without the knowledge-db still seeds the non-knowledge shots and only warns).
• ensureKnowledgeEntryDocumentsIndexed: the knowledge entries' markdown documents share the documents-list frame, so their Failed rows are retried too.
• documents-list gate waits for the last file's Indexed badge instead of its filename.
• README: TALE_DEV_SKIP_DOCKER=1 only skips docker compose up; the backend still dials the knowledge-db on :5433, so with the containers already running indexing works.

Product findings (1 and 2 fixed below; 3 not fixed here)

1. Knowledge-entry documents land at the Documents root. KNOWLEDGE_ENTRIES_FOLDER ("Knowledge entries", backend/core/knowledge_entries/constants.ts) has no usage since fd5c009 (feat(platform): move the platform off Convex onto Postgres #3107); the 0.4 rig shot showed them filed in that folder. The new documents-list shot and captions describe the 0.5 behaviour; regenerate once the folder is restored. → fixed in this PR (see Also fixes): the constant is dead code and is dropped; the root listing is the 0.5 behaviour.
2. rag_error for an interrupted index says "Re-upload the file to try again", but Retry indexing recovers it — the message could name the retry. → fixed in this PR (see Also fixes).
3. The seeder still calls bunx convex run agents/installations:upsertInstallation for the researcher agent and skips on failure (pre-existing, logged every run).

Rig state handled through the UI

• A duplicate relaunch-content-inventory.txt (seeder race from an earlier run, since fixed) was detached from the project and deleted from Knowledge.
• 2026-brand-guidelines.txt carried a v1 · Draft record badge from the earlier run's controlled-document-replace-file shot (records cannot be un-controlled); it was deleted and re-uploaded by the seeder, which also exercised the plain upload → index path.

Gates

• bun run --filter @tale/docs test — 194/194 (30 files, incl. the de/fr docs i18n checks); bun run --filter @tale/docs lint — clean.
• oxlint on the changed platform files — clean; tsc --noEmit -p services/platform/tsconfig.json — rc 0; oxfmt --check — clean.

Also fixes

Two follow-up commits close product findings 1 and 2 above:

• fix(platform): point interrupted-indexing errors at Retry indexing — the RAG reconcile sweep's rag_error for a dead ingest chain (backend/domains/file_metadata/watchdogs.ts, both settle branches) now reads "Indexing was interrupted before it finished. The stored file is intact — use Retry indexing to index it again." instead of asking for a re-upload. rag_error prose is an English backend literal rendered verbatim in the failed dialog (like its siblings in domains/knowledge/service.ts), so there is no locale work, and no rag_error_code is added because Retry indexing already sits beside every failed badge. The constant is exported: the real-Postgres probe (integration-check.ts) asserts the exact text instead of sniffing for "interrupted", and a new unit lock (domains/file_metadata/watchdogs.test.ts, vitest server) pins that the guidance names Retry indexing and never a re-upload, that the corpus's own error always wins over the generic text, and that an already-failed row is never overwritten with it.
• chore(platform): drop the unused knowledge entries folder constant — KNOWLEDGE_ENTRIES_FOLDER removed from backend/core/knowledge_entries/constants.ts. No importer since feat(platform): move the platform off Convex onto Postgres #3107; knip never flagged it because every backend/**/*.ts is a knip entry for the platform workspace. rg over services/platform and docs/{en,de,fr} finds no message key, helper, test, or docs sentence that existed only for it (the seeder's comment describing the root listing stays — it is the 0.5 behaviour).

Gates on the final tree: vitest server — file_metadata/watchdogs.test.ts, lib/i18n/messages.test.ts (orphan-key guard), core/knowledge/index_health.test.ts, core/knowledge_entries/helpers.test.ts (4 files, 69 tests); tsc --noEmit rc 0; oxlint --type-aware and oxfmt --check clean on the four touched files; knip rc 0. Docs are untouched by these two commits. The backend integration probe was type-checked, not re-run.

Draft — not for merge yet.