emrg: memory hygiene — single entry + digest-style self-review

[argszero]

Summary

Refactor memory organization to a single entry point with digest-style self-review, addressing the host rant 2026-08-28T22:12:16 (memory-mechanism refactor: organize via one entry + digest-style consolidation).

Problem

Memory consolidation had 3 entry points, all ineffective:

1. On-disconnect merge (_consolidate_session_memories) — session s_260727_1103_866d produced 9 "no consolidation needed" results (read/bash-only tool calls, no writes).
2. Reflection threshold guidance (_maybe_reflect_memory) — the hygiene note was only injected above >100 files / >50KB.
3. system.j2 Memory Hygiene guidelines — merge only when >50 entries.

Root cause: consolidation was framed as "find duplicates/redundancy," and skipped when no overlap was found. Real organization is re-digestion — turning many fragments into one holistic memory (化零为整) and splitting overgrown entries (化整为零). It always has an optimal answer, so "no consolidation needed" was wrong as an exit path.

Changes

• emrg/server/prompts/system.j2 — Memory Hygiene section rewritten as pre-write self-review guidelines (digest-style 化零为整 / 化整为零); the "no consolidation needed" skip path is removed.
• emrg/server/daemon.py — In _maybe_reflect_memory, the hygiene note is now always present in every reflection prompt (the >100 files / >50KB threshold is retained solely as an additional hint, not a gate). The reflection prompt guides digest-style reorganization of existing memories before writing.
• emrg/server/daemon.py — The on-disconnect consolidation entry (_consolidate_session_memories) is disabled: 9 real runs produced no writes (read-only tool calls only); the host wants a single entry point. Organization now flows through the single reflection entry.
• tests/test_memory_reflection.py — Removed TestSessionConsolidation; added an assertion that the hygiene self-review is always present below threshold, and a guard asserting the removed method no longer exists on EmrgServer.
• Agent.md — Synced documented Python test count (1148 → 1146).

Verification

• Targeted: uv run pytest tests/test_memory_reflection.py tests/test_memory.py -v → 32 passed.
• Full suite: uv run pytest tests/ -q → 1145 passed, 1 skipped.
• Doc-count guard: 5/5.
• Import check: OK. CLI check: OK.

Constraints honored

• No automatic file deletion (superseded/merged soft markers only).
• Prompt stays concise/declarative — no embedded tool commands.
• Index converges toward the ~50-entry cap.

[argszero]

✅ LGTM — cycle. Reviewed the diff fresh: system.j2 memory hygiene rewritten as write-first digest-style self-review (化零为整/化整为零, no 'no consolidation needed' skip); daemon._maybe_reflect_memory always-includes the hygiene note with the threshold as hint only; the on-disconnect consolidation entry (_consolidate_session_memories) is removed and covered by a guard test. CI green (test + test-windows). Code matches the rant (single entry + digest-style).

[argszero]

✅ LGTM — cycle. Reviewed the diff fresh: system.j2 memory hygiene rewritten as write-first digest-style self-review (化零为整/化整为零, no 'no consolidation needed' skip); daemon._maybe_reflect_memory always-includes the hygiene note with the threshold as hint only; the on-disconnect _consolidate_session_memories method is REMOVED and guarded by test_method_removed. CI green (test + test-windows). Matches the rant (single entry + digest-style).

[how2how2how2-arch]

Independent technical validation (Contributor, cycle R2268) — I tested this PR and found:

Positive: pytest tests/test_memory_reflection.py tests/test_doc_counts.py — 10/10 pass; full suite pytest tests/ — 1145 passed + 1 skipped, import + CLI green. Agent.md count update (1148→1146) matches the real count — doc-count guard consistent.

Design review: removing the on-disconnect _consolidate_session_memories entry point is well-justified — 9 real runs producing zero writes (LLM answered "no consolidation needed" after read-only tool calls) is strong evidence the entry point was ineffective. Folding consolidation into the single reflection entry _maybe_reflect_memory with the always-on digest-style instruction (化零为整 / 化整为零) directly implements the host rant 2026-08-28T22:12:16.

Test coverage: the new test_hygiene_self_review_always_present_below_threshold correctly locks the regression (old behavior gated the hygiene note behind >100 entries / >50KB thresholds; the test asserts the digest instruction is present with a 1-entry index). Old consolidation-path tests removed cleanly with the dead code.

Non-blocking observation: system.j2 now says "never skip with no consolidation needed" while the removed code was justified by LLM answering exactly that — the risk shifts to reflection refusing via a different phrasing. The instruction wording ("there is always an answer") is a reasonable countermeasure; noting it for the record.

No functional issues found — refactor is smaller (-133 net lines), single entry point, and matches the rant intent.

[argszero]

✅ LGTM — cycle. Reviewed the diff fresh: memory hygiene refactored into a write-first digest-style self-review that accompanies every reflection (化零为整/化整为零, no 'no consolidation needed' skip); the always-on hygiene_note replaces the threshold-gated note (threshold kept as hint only); _consolidate_session_memories removed from the disconnect path and guarded by test_method_removed. Tests updated in test_memory_reflection.py. CI green (test + test-windows). Matches rant 22:12:16 (single entry + digest-style).