emrg: memory index governance — cap index size and enforce hygiene

[argszero]

Summary

Addresses host rant 2026-08-23T08:04:26 (project emrg): MEMORY.md must stay a pure index and converge to ≤50 entries so the index can never bloat request bodies (413).

Evidence: a long-lived evolution session's memory/MEMORY.md reached 280,943 bytes / 752 lines (731 per-cycle index rows, avg 384 chars each). Root cause: _consolidate_session_memories only fires on client disconnect, so long-lived daemon sessions never trigger consolidation, and reflection was append-only.

Changes

1. emrg/memory.py
   • Write-time title truncation (≤512 chars) in MemoryIndex.add_entry (primary guard)
   • Render-time fallback in to_markdown for legacy dirty index data (filename kept reachable)
   • SessionMemoryStore soft count/size guard — logs a warning when index > 100 entries or > 50 KB (warn only, never auto-deletes; consolidation stays LLM-driven)
2. emrg/server/daemon.py
   • _maybe_reflect_memory: when the session index crosses thresholds, injects a Memory Hygiene section steering the reflection LLM toward consolidation (merge/supersede instead of append)
   • _consolidate_session_memories: adds an explicit index cap — index must converge to ≤50 entries (detail .md files remain the source of truth and may exceed 50)
3. Prompts — system.j2 (Memory Hygiene rules), evolution_prompt.md (Step 6 index-line norm), open_source_prompt.md (Recording memory-hygiene note)
4. Tests — +4: write-time truncation, render-time truncation fallback, soft-guard silent-below-threshold, soft-guard warns-at-threshold (1006 passed + 1 skipped)

Local remediation (outside this repo, host ~/scm copy, .emrg/ gitignored): the 280 KB aitokenpool session MEMORY.md was trimmed to 23,897 bytes / 50 rows; its 597 cycle-*.md detail files are untouched (source of truth).

Verification

• uv run pytest tests/ → 1006 passed, 1 skipped
• uv run python -c "from emrg.client.app import run_client" → OK
• uv run python -m emrg --help → OK

[argszero]

✅ LGTM — cycle (self-review)

Verified locally:

• uv run pytest tests/ → 1006 passed, 1 skipped (+4 new memory-governance tests)
• uv run python -c "from emrg.client.app import run_client" → OK
• uv run python -m emrg --help → OK

Changes are scoped: write-time title truncation + render-time fallback + soft guard in memory.py; hygiene steering in daemon.py reflection/consolidation prompts; Memory Hygiene rules in system.j2 and both prompt templates. Consolidation remains LLM-driven — nothing auto-deletes.

[argszero]

✅ LGTM — cycle (independent review)

Fresh-eyes review of the full diff:

• _truncate_index_title uses a None sentinel default (avoids the NameError), write-time truncation is primary, render-time fallback covers legacy dirty data — filename (detail file) preserved in both paths.
• Soft guard (SessionMemoryStore._warn_index_thresholds) is warn-only and non-destructive; consolidation stays LLM-driven via the updated reflection/consolidation prompts (index cap ≤50 entries, pure-index rule).
• Hygiene note in _reflect is computed on-demand and only appended when thresholds are crossed — no overhead in the common path.
• +4 tests cover both truncation paths and both guard states (below-threshold silent / at-threshold warns + non-destructive check); Agent.md count 1003→1007 matches.
• CI: test + test-windows both pass (run 32607536232).

No issues found.

[argszero]

✅ LGTM — cycle (3rd consecutive approval)

Head 8de48f6 unchanged since the previous two reviews; CI test + test-windows both pass (run 32607536232). Full diff was reviewed fresh in the previous cycle: write-time + render-time title truncation with filename preservation, non-destructive soft guards, LLM-driven consolidation steering, +4 tests matching the Agent.md count bump (1003→1007). No issues — approving to merge.