[schemas] enhanced-thoughts: harden upsert/match/search RPCs

[alanshurafa]

What this does

The enhanced-thoughts schema shipped an older generation of the upsert_thought and search_thoughts_text RPCs than the reference brain has been running. This PR closes that gap. Everything is additive and idempotent — re-running schema.sql on a v1 install is a no-op, and the v1 return contract ({id, fingerprint}) plus the status / status_updated_at handling that workflow-status depends on are left exactly as they were.

Changes

• search_thoughts_text learns date and tier filters. p_filter now reads three reserved control keys — start_date, end_date (ISO 8601), and exclude_restricted (boolean) — and applies them at the data layer. They are peeled off before the metadata @> filter containment check, so any other key keeps its old behavior.

• upsert_thought gets two dedup/merge guards.

  • Original-fingerprint fallback. When a thought's content is corrected its fingerprint changes; a later reimport of the original text used to insert a stale sibling that outvoted the correction. If the pre-edit fingerprint lives in an append-only metadata.original_fingerprints[] array, the reimport now lands on the corrected row instead.
  • User-edit guard. Keys listed in metadata.user_edits are treated as human-owned and stripped from an incoming automated patch, so a reimport can't clobber a human correction. Guards metadata keys only.
  • To make the fallback possible the function does an explicit lookup and branches INSERT vs UPDATE instead of ON CONFLICT. The unique-index race that ON CONFLICT handled for free is caught explicitly and folded back into the merge path.

• New opt-in match_thoughts_superseded_aware. Same shape as the core match_thoughts plus a superseded_by column. Thoughts that have been replaced (the target of a supersedes edge in thought_edges) take a 0.8x ranking penalty so fresh thoughts surface above their stale predecessors, without ever being excluded. The core match_thoughts is untouched; callers opt in by name, matching the recency-boosted-match-thoughts pattern. Installed only when schemas/typed-reasoning-edges/ is present; otherwise it's skipped with a NOTICE and the rest of the migration still applies.

ID contract

All ported references use UUID against public.thoughts(id). The superseded-aware RPC returns superseded_by UUID and reads thought_edges where from_thought_id is the newer replacement and to_thought_id is the stale thought — the directional semantics OB1's typed-reasoning-edges already documents.

Importance scale

Some installs use a narrower 0-6 importance scale. Open Brain's upsert already accepts a wider 0-100 range, so it never clipped 0-6 values — switching to 0-6 here would retroactively rescale every existing row, which is a breaking data change, not an additive one. The 0-100 clamp stays; the README documents 0-6 as a subset for cross-scale parity.

Testing

Applied against a throwaway PostgreSQL instance with stub thoughts + thought_edges tables:

• Insert returns {id, fingerprint}; re-insert dedupes to the same row.
• Original-fingerprint reimport lands on the corrected row (no stale sibling; content preserved).
• User-edit guard keeps a human title against a robot overwrite.
• start_date and exclude_restricted filters return the expected subset.
• match_thoughts_superseded_aware returns the stale row at 0.8x with superseded_by set, fresh row at full similarity.
• Re-running schema.sql is clean; dropping thought_edges triggers the documented skip-with-NOTICE.

Compatibility

Preserves the workflow-status dependency documented in #328 — upsert_thought still writes status / status_updated_at, and the README's Prerequisites now call out workflow-status explicitly.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c6f972b990

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve corrected rows after original-fingerprint matches

When matched_via is original_fingerprint, this guard skips only embedding, but the subsequent update still writes metadata, promoted fields, and status derived from the stale original capture. That bypasses the RPC's user_edits / original_fingerprints merge protection and can overwrite the corrected row's human-owned metadata or workflow state whenever a REST reimport hits the fallback; in this branch the update should be skipped or limited to fields that cannot be stale.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Preserve existing task statuses on duplicate upserts

On merge paths where the caller omits status, this re-derives v_status = 'new' from the effective existing type, so any duplicate or metadata-only re-upsert of a task/idea that is already planning, active, done, etc. is reset back to new and gets a fresh status_updated_at. That is a regression from preserving omitted fields/status; only seed new on insert or when the existing row's status is null.

Useful? React with 👍 / 👎.