Recall correctness debt discovered during memory R&D

[heybeaux]

Keep these defects separate from the memory-formation/query-transformation ablation so they do not muddy causal attribution.

1. Flat identity-rescue scoring

src/memory/memory-query.service.ts injects identity-rescue rows at a flat 1.15, inside the FTS rescue band. Preserve relevance evidence, define deterministic ordering, and add non-domination tests.

2. Unreachable forced-FTS fallback

The forced-FTS branch assigns 0.75, but current topIds/memoryMap construction appears to make it unreachable. Prove reachability, then remove or repair it with a regression test.

3. Embedding model/dimension mismatch

Research instrumentation observed 1536-d OpenAI vectors meeting 384-d/local-model storage expectations on the write path. Trace model identity end-to-end, reject/route mismatches before SQL, and cover 384/768/1536 providers mechanically.

4. Local embedder saturation

Bulk ingest (about 3.5k chunks in parallel) has produced malformed sparse vectors like [,,,,,...], causing pgvector 22P02 and memories without embeddings. Add backpressure, finite-vector validation, explicit metrics, and rerun the LongMemEval/bulk-ingest reproducer. Preserve the OpenAI-embedding workaround until verified.

Constraints

• No production data mutation.
• Do not combine these fixes with formation/query-rewrite experiments.
• Measure recall and downstream behavior independently.