fix: /recall entity/category post-filter is inefficient

[ajianaz]

Summary

/recall entity/category filters (main.rs:732-759) are post-filters in Rust AFTER semantic search. This means:

1. Fetches up to limit*10 results (max 200)
2. Rust iterates all results checking metadata
3. Non-matching results discarded

A memory matching entity filter perfectly might be excluded if not in top 200 by similarity.

Proposed Fix

Push entity/category filter into SQL query layer before scoring.

Acceptance Criteria

Functional

• Entity/category filter applied at SQL level, not Rust post-filter
• No 10x fetch amplification
• Correct results for low-similarity but matching memories

Verification (CMO can run without reading code)

• curl POST /recall {"query":"launch","entity":"drawover"} → returns drawover memories even if embedding similarity is low
• curl POST /recall {"query":"launch","entity":"nonexistent"} → returns empty results (no entity match)
• Compare POST /recall {"query":"generic","entity":"drawover"} results before and after fix → same or more accurate results
• Recall time with entity filter < 50ms (vs current ~200ms with 10x amplification)

Edge cases

• Entity filter on memory with no entity metadata → memory excluded (correct behavior)
• Multiple memories same entity → all returned (not just top-1)

Tests

• Unit: entity filter at SQL level returns correct subset
• Benchmark: entity-filtered recall < 50ms

Priority

Low — correctness issue for edge cases.

Dependencies

Partially addressed by #655 (metadata storage) and #656 (DB indexes).

[ajianaz]

Dependencies

• feat: add custom metadata fields to RememberRequest API #655 (custom metadata storage)
• fix: /remember does not set memory_type column #656 (DB indexes)

Alternative

If #658 (POST /query) lands first, entity/category can be handled via structured queries instead.