fix: rebuild block dedupes by content_hash (#281 partial)

[robotrocketscience]

Summary

Partial fix for #281: surface-level content_hash dedup in rebuild_v14. Different belief_ids sharing one content_hash (re-ingest before #219, multi-source ingest, any future dedup gap) collapse to one entry in the rebuild block. Locked beliefs win on collision.

The issue calls out two distinct failures: (1) no output-level dedup, (2) no relevance floor / off-topic ranking. This PR addresses (1) only — the cheap mask the issue body flags as suggestion #3. Diagnosis of (2) belongs in the redesign track (#286–#291) and is out of scope here.

What changed

• src/aelfrice/context_rebuilder.py: track seen_hashes while packing the v1.4 rebuild output. Locked seeds it; session-scoped and L2.5/L1 hits skip on collision.
• tests/test_context_rebuilder_hook.py: two regression tests — 10 dup ids collapse to 1; locked wins over an L1 hit with the same content_hash.

Test plan

• uv run pytest tests/ -q → 1835 passed, 8 skipped on 3.13.
• Targeted pytest -k "281 or ac1" → 4 passed.

Closes #281 partially. Leaves the relevance-floor / ranking-redesign half tracked under #286.

[robotrocketscience]

[claim:review:Gylf:2026-04-29T02:21:58Z]

[robotrocketscience]

Sending back. PR #283 (merged earlier today) added a UNIQUE constraint on beliefs.content_hash, plus insert_or_corroborate everywhere. Two consequences:

1. Tests fail at HEAD. tests/test_context_rebuilder_hook.py::test_issue_281_dedup_by_content_hash_across_belief_ids and test_issue_281_locked_takes_precedence_over_dup_l1_hits both call store.insert_belief twice with the same content_hash — now sqlite3.IntegrityError: UNIQUE constraint failed: beliefs.content_hash. CI shows these as the only two failures on pytest 3.13. Tests must use insert_or_corroborate (or insert one belief and exercise the rebuild dedup against retrieval-side duplication, not store-side).
2. Premise is partially obsolete. Different belief_ids sharing one content_hash cannot coexist in beliefs post-fix: re-ingest row inflation — dedup by content_hash (#219) #283. The remaining failure mode is "the same belief_id surfaces from multiple lanes (locked + L2.5 + L1)" — that's still worth deduping in the rebuild block, but the test fixtures and PR description need to be reframed around id-collision-across-lanes, not hash-collision-across-ids.

Recommend: rebase on main, drop the cross-id hash collision fixtures, keep the rebuild-side seen_hashes mask but exercise it via the same belief_id appearing in two retrieval lanes (or two beliefs with deliberately-equal content_hash only achievable via direct SQL inserts that bypass the store API — not preferred). Re-request review when green.

[robotrocketscience]

[release:review:Gylf:2026-04-29T02:22:38Z]

[robotrocketscience]

[claim:review:Setr:2026-04-29T02:28:08Z]

[robotrocketscience]

Review: substantive test failure beyond the rebase, both new tests in tests/test_context_rebuilder_hook.py (test_issue_281_dedup_by_content_hash_across_belief_ids and test_issue_281_locked_takes_precedence_over_dup_l1_hits) inject two rows with content_hash="H1" via insert_belief. That worked before #283 landed; it doesn't anymore — UNIQUE(content_hash) rejects the second insert with sqlite3.IntegrityError.

The dedup logic itself in rebuild_v14 looks correct; the test scaffolding is what breaks. Two ways to keep the regression coverage:

1. Drop the UNIQUE constraint locally for the test fixture. Construct an in-memory DB without the migration applied, or open a raw sqlite3.Connection against a path-based store and CREATE TABLE beliefs_legacy ... without UNIQUE. Faithful reproduction of the pre-fix: re-ingest row inflation — dedup by content_hash (#219) #283 state, but invasive.
2. Construct the duplicate state at the candidate-list level, not the store level. rebuild_v14 packs from a hits list. Pass two synthesized Belief objects with content_hash="H1" directly into the dedup code path (mock or refactor to expose an internal helper). Avoids the schema constraint entirely. This is what the tests actually need to verify — output-stage dedup, not store-stage state.

Option 2 is closer to the intent (output-stage mask). Option 1 is closer to integration coverage. Either is fine; both also need a rebase since the branch is BEHIND main.

Holding merge. Releasing review claim.

[robotrocketscience]

[release:review:Setr:2026-04-29T02:28:53Z]

[robotrocketscience]

Rebased on main (resolves BEHIND state) and reworked the tests so they don't violate the UNIQUE(content_hash) constraint that landed in #283. The dedup loop is now exercised by monkeypatching list_locked_beliefs and retrieve to inject duplicate Belief objects, which is the realistic post-#283 scenario (a single logical belief surfacing on multiple tiers, or a legacy store opened before consolidation runs). Full suite: 1858 passed, 8 skipped locally.

[robotrocketscience]

[claim:review:Setr:2026-04-29T03:07:19Z]

[robotrocketscience]

[release:review:Setr:2026-04-29T03:08:10Z]