[integrations] Consolidation workers (bio + metadata)

[alanshurafa]

Summary

Two Deno/Supabase Edge Functions for post-import thought quality improvement:

• `metadata-norm` — re-classifies the `type` and `importance` fields on accumulated thoughts that were captured pre-enrichment or with weak metadata
• `bio` — synthesizes a canonical "Who is {subject}" profile thought from biographical captures scoped to that subject

Both workers use three-tier LLM fallback (OpenRouter → OpenAI → Anthropic) with transient-error gating (`isTransientError` predicate) so provider hops only fire on 5xx/429/timeout — not on 4xx/auth errors that would waste calls.

Hardened per Wave 2.5 review:

• `CONSOLIDATION_MAX_CALLS` budget (default 100, 0 = unlimited); preserves `consolidation_reviewed` markers on trip
• AbortController on every LLM fetch (`FETCH_TIMEOUT_MS` default 60s)
• Prompt-injection defense: thought content wrapped in `<thought_content>` tags (with closing-tag escape via zero-width-space); `<previous_profile>` wrapping in bio; system/user role split on all 6 provider calls; `importance>=6` output treated as an injection signal (clamps to 0-5)
• Bio insert via `.insert()` with `content_fingerprint` instead of `upsert_thought` RPC (fixes the RPC-signature mismatch that dropped `type`/`importance`/`source_type` on first-run profiles)
• Per-subject profile scoping: `findExistingProfile` filters on `metadata->>subject` — profiling multiple subjects no longer overwrites each other's rows
• UUID types for `thought_id` throughout; dead `typeof === 'number'` branches removed
• b361df1 name-filter applied across all three source queries in the bio worker

Why

Open Brain's import paths accumulate thoughts faster than classification can keep up, especially for pre-enrichment backfills (drag-and-drop archives, Takeout dumps). `metadata-norm` is a targeted re-classifier for the long tail. `bio` is a different shape: synthesize a stable "self" or "about-{person}" anchor document that LLMs can efficiently ground against.

Test plan

• POST `/metadata-norm` with `{ "limit": 5 }` — verify 5 thoughts re-classified, `consolidation_reviewed: true` set
• Set `CONSOLIDATION_MAX_CALLS=2`, post with `limit=5` — verify 2 processed, loop breaks, response includes `truncated: { reason: "budget_reached" }`
• POST `/bio` with `{ "subject": "self" }` on first run — verify a profile thought is created with correct type/importance/source_type (not null)
• POST `/bio` twice for `subject: "Sarah"` then `subject: "Alice"` — verify Sarah's profile is not overwritten by Alice's
• Kill primary LLM provider (stall), verify fallback fires through to secondary
• Post with hostile content (prompt-injection attempt) — verify extraction returns clean values, `importance` stays 0-5
• `deno check` on both `index.ts` files + `_shared/`

[alanshurafa]

Conflicts with main. Before I rebase, should this stay as-is, be reshaped, or be dropped? I can refresh it quickly if you want to keep it.

[alanshurafa]

Rebased onto main — conflicts cleared. The only conflict was a stray repo-wide markdown-lint commit on this branch that collided with main's own lint cleanup; I dropped that commit. The branch's actual changes are untouched. Now mergeable.

[justfinethanku]

Validated both worker entrypoints with deno check and added the missing OB1 community credit header. Ready to merge.