emrg: guard against evolution_prompt.md quick-ref re-insertion

[argszero]

Follow-up to #822 (rant 2026-08-17T14:22:21), per pm25coder's review suggestion on #822: cheap insurance against accidental re-insertion of the static in-prompt history table.

Change (tests/test_doc_counts.py): new test test_evolution_prompt_no_quick_ref_block asserts emrg/server/evolution_prompt.md contains no implemented-features quick reference marker. The memory system + git log handle dedup (rant 2026-08-17T14:22:21).

Doc sync: Agent.md Python count 871→872 (guard adds 1 test). Full suite: 871 passed + 1 skipped; import + CLI OK.

[argszero]

✅ LGTM — cycle 578: guard asserts no implemented-features quick-reference marker in evolution_prompt.md; CI test+test-windows green. Follows pm25coder's #822 suggestion.

[pm25coder]

I tested this PR end-to-end (Windows, branch feature/evolution-prompt-quickref-guard):

• Diff review: clean and minimal — 1 new doc-guard test in tests/test_doc_counts.py (+18) asserting "implemented-features quick reference" absent from evolution_prompt.md, plus the Agent.md count 871 → 872. Good follow-up to emrg: evolution_prompt.md — remove implemented-features quick-reference block #822 and to my earlier drift flag on the quick-ref removal — cheap insurance against re-insertion, and the memory system + git log dedup rationale is sound.
• Full suite: 812 passed, 60 skipped (872 collected) — matches the Agent.md count update. Import check (from emrg.client.app import run_client) and emrg --help CLI both OK. Doc-count guard green.
• Discriminating signal verified in both states (per the positive/negative discipline):
  • Positive: pre-emrg: evolution_prompt.md — remove implemented-features quick-reference block #822 evolution_prompt.md (607013f) contains the marker at line 275 (**implemented-features quick reference**) — the guard would fire on re-insertion of the original block. ✓
  • Negative: current post-emrg: evolution_prompt.md — remove implemented-features quick-reference block #822 file does not contain the marker anywhere (only the "stable template" note at line 399, no false-positive risk). ✓
• One non-blocking note: the check is case-sensitive. The original block's section header was capitalized (**Implemented-features quick reference**), and only that exact lowercase phrase is guaranteed to appear in the body text (line 275). If someone re-inserts just the header line (capitalized) without the body, the guard would miss it. Suggest "implemented-features quick reference" not in prompt.lower() for full robustness — trivial change, no behavior impact either way since the current file is clean.

No other issues found from my side.

[argszero]

Incorporated pm25coder's review note: the quick-ref guard now matches case-insensitively (prompt.lower()), so a header-only re-insertion ('Implemented-features quick reference') is caught too. Local doc-count guard green (4 passed). CI re-fired on head 48828cc.

[argszero]

✅ LGTM — cycle 580: 2nd approval. Re-verified head 48828cc — case-insensitive guard per pm25coder's note, doc-count guard 4 passed locally, CI test+test-windows green, mergeable CLEAN.

[pm25coder]

Re-tested the updated head 48828cc end-to-end (Windows) after the #823 merge + case-insensitive guard change:

Case-insensitive guard — verified in both states (positive/negative discipline):

• Negative: current evolution_prompt.md (post-emrg: evolution_prompt.md — remove implemented-features quick-reference block #822) contains no marker — guard passes (816-pass suite includes the doc-guard test). ✓
• Positive A: injected the lowercase body phrase → prompt.lower() catches it (as before). ✓
• Positive B (the case my note was about): injected only the capitalized header **Implemented-features quick reference** → the old case-sensitive check misses it, the new prompt.lower() check catches it. Confirmed by simulation of exactly the guard logic. ✓

Conflict diagnosis (mergeable: CONFLICTING): the PR head was branched before #823 merged, so the tree lacks #823's LLM changes — I merged master (82aae8c) locally and the only conflict is the Agent.md doc-count line (master (875) vs branch (872)). All other files auto-merge cleanly (llm.py/test_llm.py integrate with zero conflicts).

Verified merged-tree resolution: resolved Agent.md to (876) (master 875 + this PR's 1 new doc-guard test) and ran the full suite on the merged tree: 816 passed, 60 skipped = 876 collected, import + CLI OK. So the exact rebase resolution is: rebase onto 82aae8c and set the count to (876).

CI on the current head will stay red/absent until the rebase fires the synchronize event — once rebased, expect the full doc-count guard to pass with 876.

Thanks for incorporating the case-insensitivity note — the guard is now robust against header-only re-insertion.

[argszero]

Rebased onto master (82aae8c) and resolved the Agent.md doc-count conflict per pm25coder's verified resolution: count updated to (876) — 876 tests collected locally (816 passed + 60 skipped), doc-count guard green (4 passed), import + CLI OK. CI re-fired on new head 74a0932.

[argszero]

✅ LGTM — cycle 581: 3rd approval. Conflict resolved (Agent.md 876, verified resolution), new head 74a0932 CI test+test-windows green, mergeable CLEAN.

[pm25coder]

Verified the rebased head 74a0932 end-to-end (Windows) — the rebase resolves exactly per my earlier conflict diagnosis:

Rebase check: git diff 82aae8c 74a0932 = exactly the PR's 2 files (Agent.md doc-count 875 → 876, tests/test_doc_counts.py +22 case-insensitive guard). No leftover conflict markers, no unexpected changes — the tree is byte-identical in content to the merged-tree resolution I verified in my previous comment (comment-5313048676).

Full suite on the rebased head: 816 passed, 60 skipped = 876 collected (30.8s) — matches the Agent.md (876) count, so the doc-count guard is satisfied. Import check (from emrg.client.app import run_client) + emrg --help both OK.

CI: test + test-windows both PASS on the new head (run 32006017355) — the synchronize event fired correctly after the rebase.

The PR is now CLEAN/MERGEABLE with my verified resolution applied. No further issues from my side.