Prioritize chat over memory synthesis

[joelteply]

Summary

• default Hippocampus consolidation to raw memory so background memory work does not consume the visible chat inference lane
• keep semantic compression available only behind CONTINUUM_ENABLE_LLM_MEMORY_SYNTHESIS
• move Hippocampus to the lowest background cadence, pause it during startup autonomous-work gating, and require strict background backpressure for opt-in semantic synthesis

Validation

• npx vitest run tests/unit/memory/HippocampusConsolidationPolicy.test.ts tests/unit/memory/RawMemoryAdapter.test.ts tests/unit/memory/SemanticCompressionAdapter.test.ts --no-color
• npx tsc --noEmit --project tsconfig.json --pretty false
• precommit: TypeScript build, modified-file ESLint baseline check, browser ping
• prepush: TypeScript clean, ESLint baseline-tolerant clean

[joelteply]

LGTM — small surgical fix that targets exactly the 'Compressed Insight' canned-text leak we observed when memory consolidation was sharing the visible-chat inference engine.

Verified

1. Default consolidation = RAW (no LLM): selectDefaultConsolidationAdapter reads getDefaultConsolidationMode() which defaults to 'raw' unless CONTINUUM_ENABLE_LLM_MEMORY_SYNTHESIS=1. RawMemoryAdapter is a real pass-through (WorkingMemoryEntry → MemoryEntity, no synthesis call) — memory still gets stored, just without the LLM step that was bleeding into chat replies.

2. Priority 'low' → 'lowest': doc says "Background memory must not compete with visible chat turns." Belt-and-suspenders.

3. Hippocampus.tick short-circuits on StartupAutonomousWorkGate.isPaused(): same gate as Stabilize startup persona backpressure #1058. Memory consolidation deferred during seed/cold-start, consistent with the autonomous-loop pattern.

4. SemanticCompressionAdapter BackpressureService check upgraded 'low' → 'background': when LLM synthesis IS opted in, it now uses the strictest back-off lane.

5. Test: covers env-var policy function (default raw, opt-in semantic). 2 cases.

Non-blocking observations

• Test coverage gap: the policy function is tested but the larger behavior changes — RawMemoryAdapter selection at construction, 'lowest' priority, StartupAutonomousWorkGate.isPaused() short-circuit, 'background' backpressure level — aren't directly asserted in tests. The first one in particular is the user-visible behavior change. A simple Hippocampus constructor test (env unset → adapter is RawMemoryAdapter; env=1 → SemanticCompressionAdapter) would lock the wiring in. Worth a follow-up.

• 'lowest' priority value: assuming the PersonaContinuousSubprocess priority enum supports it (TS would catch otherwise — the prepush passed). Worth checking the enum has lowest distinctly below low so this isn't a no-op rename.

• Documentation: when CONTINUUM_ENABLE_LLM_MEMORY_SYNTHESIS=1, what's the user-facing behavior difference? Briefly: (a) memory entries become semantic summaries instead of raw thoughts, (b) consumes inference budget that could otherwise serve chat. Worth a one-line operator note in the env var docs.

LGTM ship.