fix(sleep): make epistemic promotion reachable — namespace spread never fired

[salishforge]

The defect

Phase 5.12 promotes a memory to established only when it has positive retrievals from ≥ 2 distinct namespaces. That can never happen:

• every search path filters warm rows to the caller's namespace (memory-manager.ts:975, 1015, 1052, 1092)
• retrieval_log records that same namespace
• namespace is immutable after insert — no SET namespace anywhere in src/

A row's retrieval evidence is therefore always single-namespace, so COUNT(DISTINCT namespace) cannot reach 2.

Nothing has ever been promoted to established in production. That also silently starved bootstrapAgent(), which copies only established rows — so cross-agent transfer learning moves zero memories. Two shipped features, both inert.

Both had passing tests, because the suite hand-inserted retrieval_log rows carrying a namespace the system cannot produce for that row.

The fix

Replace the namespace clause with a day-spread test: positive retrievals on ≥ 2 distinct calendar days. Preserves the documented intent (corroboration from independent sessions rather than one burst) and is reachable.

Not redundant with evidence_count >= 3: that counter increments once per sleep cycle that saw feedback in the trailing 24h, so several cycles in one busy day can reach 3 without any independent confirmation. Both clauses now do distinct work.

Tests

• Reachability guard — drives promotion through the real query() + feedback() paths, backdating only a timestamp. If promotion ever becomes unreachable again, this fails. (Its poll is anchored to a pre-call id watermark because retrieval logging is fire-and-forget — memory-manager.ts:942.)
• Invariant pin — asserts no retrieval row carries a namespace differing from its memory, so a future reintroduction of namespace-spread fails loudly with the reason.
• Flake fix — promotion tests now each run on their own agent. They all call engine.run(), which mutates every eligible row for that agent, so the shared agent let a later cycle retroactively promote an earlier test's fixture. Reproduced at ~1-in-4, then absent across 8 consecutive runs.

Suite	Result
test:epistemic-confidence	33/33 (×8 runs)
test:bootstrap	22/22
test:contested-conflicts	9/9
test:integration	24/24
type-check / lint	clean

Note

This is step 1 of the profiles-over-namespaces direction: it unblocks established status independent of which scoping model wins, and removes the last load-bearing use of namespace as a semantic signal.

🤖 Generated with Claude Code

https://claude.ai/code/session_011xCqQo49d3CEbn6oEvb3Ru