[idapixl]

Why

On the SQLite and JSON stores, recall always answered count: 0 (#122) and contradict answered "not found" for every observation_id from observe/wonder/speculate and every belief_id from believe (#123). Both tools read through the generic collection API (query/get on 'observations'/'beliefs'), but those writers use the typed tables. On SQLite and JSON these are two separate places to store rows. Firestore keeps one collection per kind, so the split never showed there.

The same generic read also left these unable to see any typed observation: predict's recent context, ruminate's recent observations, the PE-saturation metric, and the health, report, anomalies and vitals CLI commands.

A second bug sat underneath. The generic query() range operators compared numbers only, so every date filter on SQLite/JSON (updated_at >= <Date> on threads, traces, …) matched nothing.

What

• The store gets typed reads: getObservation(id), queryObservations({ since, until, content_type, processed, has_prediction_error, order, limit }) and getBelief(id). They follow the precedent signals already set, including the fallback: SQLite and JSON also return rows written generically, meaning older notice observations and Firestore-migrated belief revisions. getBeliefHistory merges the legacy belief revisions too. A NULL content_type reads as declarative.
• All the callers listed above move to the typed read.
• notice writes through putObservation, which leaves one observations table. Noticed rows now enter consolidation on the local stores, as they already did on Firestore; the create phase embeds them.
• Generic range filters compare dates as timestamps. Mixed Date, ISO string and serialized Firestore Timestamp values all compare correctly (shared helper stores/_generic.ts).

Verification

• New test file src/stores/typed-observation-reads.test.ts (16 tests, runs on both SQLite and JSON). Its four tool-level repros fail against master's recall/contradict/notice and pass with this change.
• Full suite: 447/447. tsc clean, build clean.
• I ran the built engine against a copy of a real 3,226-observation SQLite store:
  • queryObservations returns all 3,226 rows (recall returned 0 before).
  • Interrogative filtering returns the newest questions.
  • A legacy Firestore-era belief resolves by id.
  • One concept's history now includes 30 migrated revisions.
  • A generic threads date filter returns 16 rows (0 before).

Not in this PR

fozikio memory maintain --collection observations still reads and updates observations through the generic API, so it needs a typed update. vitals still queries edges generically. Both are filed separately.

Closes #122
Closes #123

🤖 Generated with Claude Code