fix: make the test suite Windows-green and restore full Windows support

[dfrostar]

Closes the four Windows failure classes from the cross-platform CI run (509 passed / 5 failed / 134 errors on windows-latest) and re-adds windows-latest to the gating matrix.

The four fixes

1. ChromaDB temp-dir teardown (~134 errors)

Chroma caches one System per storage path for the life of the process, holding the sqlite connection pool and HNSW segment files. GraphEmbedder.close() previously called delete_collection — which released nothing (and destroyed data a later open expected to find). It now stops the client's System and evicts it from SharedSystemClient._identifier_to_system, so Windows can actually delete the store afterwards. Verified against chromadb 1.5.9.

On top of that:

• NeuralMind.close() added (delegates to the backend, safe to call twice).
• tests/conftest.py gains an autouse fixture that stops every cached chroma System after each test — handles are released regardless of whether the test cleaned up — and the temp_project/empty_project fixtures use ignore_cleanup_errors=True as a belt-and-suspenders.

2. Event-log rotation (3 failures)

The tailer holds its read handle across poll intervals, and rotation is a logrotate-style rename. POSIX allows renaming an open file; Windows' open() omits FILE_SHARE_DELETE, so the rotating process got PermissionError. The tailer now opens via CreateFileW with share-delete on Windows, recreating POSIX semantics — rotation never depends on catching the tailer between polls. POSIX path is the same open(path, "rb") as before.

3. Concurrent recent-queries appends (1 failure)

The append relied on POSIX O_APPEND atomicity; Windows' CRT implements append as a separate seek-to-end + write, so 8 threads × 5 appends landed 37/40 lines. Appends are now a single os.write on an O_APPEND fd serialized by a process-local lock, plus a best-effort cross-process byte-range lock (msvcrt.locking, non-blocking with ~50ms retry) shared with _compact_recent_queries — so a compaction's read-truncate-rewrite can't drop a concurrent process's append either. POSIX behavior unchanged.

4. Executable-bit test (1 failure)

test_cmd_init_hook_makes_executable is skipped on Windows, which has no POSIX execute bit.

Support claims restored (issue checklist)

• Items 1–4 fixed
• windows-latest re-added to the test matrix in ci.yml (Python 3.12) — the run on this PR is the proof
• docs/COMPATIBILITY.md Windows row restored to ✅ Full
• docs/index.html operatingSystem restored to "Linux, macOS, Windows"

Also folds in the post-release landing-page touch-up (same file as the operatingSystem edit): v0.24.0 is marked as the latest release in the hero badge, timeline, and JSON-LD softwareVersion, mirroring what #224 did for v0.23.0.

Verification

• Linux: full suite 750 passed, 17 skipped; black --check and ruff check clean; mypy introduces no new errors in the touched modules.
• Windows: validated by this PR's own windows-latest CI leg (the whole point of re-gating it).

Fixes #186

https://claude.ai/code/session_01FkHXHcjpWZL2EWn4HGi547

---

Generated by Claude Code

[github-actions]

Backend parity gate — graphify vs built-in tree-sitter

✅ PASS — the built-in backend must stay within tolerance of graphify on the reference fixture.

Metric	graphify	built-in
code nodes	65	79
mean reduction	6.05×	6.66×
faithfulness delta	-0.047	+0.143
fact recall	0.527	0.717
grounding	0.889	1.000

Gate checks

• ✅ reduction within tolerance of graphify — built-in 6.66× ≥ 4.54× (graphify 6.05× − 25%)
• ✅ reduction ≥ absolute floor — built-in 6.66× ≥ floor 4.00×
• ✅ faithfulness delta within tolerance of graphify — built-in +0.143 ≥ -0.147 (graphify -0.047 − 0.10)
• ✅ faithfulness delta ≥ absolute floor — built-in +0.143 ≥ floor +0.000
• ✅ fact recall within tolerance of graphify — built-in 0.717 ≥ 0.427 (graphify 0.527 − 0.10)

Tolerances: reduction within 25% (floor 4.0×), faithfulness within 0.10 (floor +0.00). Override via NEURALMIND_PARITY_* env vars.

Automated by evals/parity/run.py — reproduce locally with python -m evals.parity.run.

Multi-language structural parity

Language	graphify symbols	built-in covers	dangling
typescript	54	54 (100%)	0
go	45	45 (100%)	0

• ✅ typescript: symbol coverage ≥ floor — 54/54 graphify symbols (100%) ≥ 90%
• ✅ typescript: no dangling edges — 0 dangling edge(s)
• ✅ go: symbol coverage ≥ floor — 45/45 graphify symbols (100%) ≥ 90%
• ✅ go: no dangling edges — 0 dangling edge(s)

Coverage floor: 90% of graphify's per-language symbols (no gold-fact set exists for TS/Go, so parity is structural).

Optional SCIP precision pass

• ✅ precision: SCIP corrects the heuristic call edge — run() → A.handle under SCIP (heuristic wrongly linked B.handle)
• ✅ precision: strict no-op when disabled — graph unchanged when NEURALMIND_PRECISION is unset

Off by default (NEURALMIND_PRECISION); proven on tests/fixtures/scip_precision to replace a heuristic call edge with the compiler-accurate one a SCIP index resolves.

[github-actions]

NeuralMind self-benchmark

Status: PASS — floor 4×, measured 6.2×.

Phase 1 — Reduction on committed fixture

• Average reduction: 6.2×
• Top-k retrieval hit rate: 71.7%
• Naive baseline: 47,360 tokens (all fixture files concatenated)
• NeuralMind total: 7,706 tokens across 10 queries
• Estimated monthly savings @ 100 queries/day on Claude 3.5 Sonnet: ~$35.69

#	Query	Shape	Naive	NeuralMind	Ratio	Hit
1	auth-flow	cross-file	4,736	773	6.1×	33.3%
2	api-endpoints	focused	4,736	758	6.2×	100.0%
3	billing-flow	cross-file	4,736	774	6.1×	33.3%
4	user-storage	cross-file	4,736	651	7.3×	50.0%
5	jwt-verify	focused	4,736	669	7.1×	100.0%
6	stripe-webhook	focused	4,736	801	5.9×	100.0%
7	create-user	cross-file	4,736	771	6.1×	50.0%
8	refund	focused	4,736	760	6.2×	100.0%
9	db-choice	identity	4,736	854	5.5×	100.0%
10	invoice-send	cross-file	4,736	895	5.3×	50.0%

Phase 2 — Learning uplift

• Memory events logged: 20
• Learned patterns: 16
• Reduction ratio after neuralmind learn: 6.1× (Δ -0.08× vs. cold)
• Top-k hit rate after learning: 75.0% (Δ +3.3 points vs. cold)

Note: uplift numbers on a 500-line fixture are intentionally modest — the point is to
verify the learning mechanism persists and applies. On real production repos the lift
is larger; this test only catches regressions in persistence.

Phase 3 — Synapse recall A/B (same warm graph, recall off vs on)

• Synapse edges after seeding co-editing sessions: 2834
• Top-k hit rate: 71.7% off → 83.3% on (Δ +11.7 points)
• Reduction ratio: 6.2× off → 6.2× on (Δ -0.07× — budget-neutral by design)

This isolates the Hebbian synapse layer from the learned_patterns reranker in
Phase 2. The hit-rate delta shows associative recall surfacing co-edited modules a
purely textual search ranks lower; the near-zero reduction delta confirms it does so
without spending extra tokens (recalled nodes displace the weakest hits, not add to them).

Assumptions

• Baseline: every .py file in tests/fixtures/sample_project/ concatenated.
• Tokenizer: tiktoken GPT-4o encoding (per-model breakdown in multi_model.json if generated).
• Pricing: Claude 3.5 Sonnet input @ $3.0/MTok.
• Regression floor: 4× — well below NeuralMind's typical 40–70× on real repos.

Per-model token reduction

Model	Tokenizer	Naive	NeuralMind	Ratio	Source
GPT-4o / GPT-4o-mini	tiktoken o200k_base	4,739	779	6.1×	measured
GPT-4 / GPT-3.5-turbo	tiktoken cl100k_base	4,710	770	6.1×	measured
Claude 3.5 Sonnet	estimated: GPT-4o × 1.08 — install anthropic for an exact count	5,118	841	6.1×	estimated
Llama 3 (70B)	estimated: GPT-4o × 1.22 — Llama tokenizer requires model weights; estimate based on published vocab ratios	5,781	950	6.1×	estimated

Rows marked measured use the provider's real tokenizer. Rows marked
estimated apply a published vocab-size correction to the GPT-4o count —
honest approximations, not hardcoded claims.

NeuralMind retrieval-quality eval

Suite	Queries	MRR	Answerability	Recall@5	Precision@5	Gate
go	10	0.950	100%	0.833	0.603	PASS
python	10	0.950	100%	0.833	0.678	PASS
typescript	10	0.900	100%	0.800	0.562	PASS

go vs baseline:

• mrr: 0.950 (= +0.000)
• answerability: 1.000 (= +0.000)
• recall@1: 0.617 (= -0.000)
• recall@3: 0.833 (= +0.000)
• recall@5: 0.833 (= +0.000)

python vs baseline:

• mrr: 0.950 (▲ +0.050)
• answerability: 1.000 (= +0.000)
• recall@1: 0.617 (▲ +0.100)
• recall@3: 0.833 (= +0.000)
• recall@5: 0.833 (= +0.000)

typescript vs baseline:

• mrr: 0.900 (= +0.000)
• answerability: 1.000 (= +0.000)
• recall@1: 0.583 (= +0.000)
• recall@3: 0.800 (= +0.000)
• recall@5: 0.800 (= +0.000)

Overall: PASS

---

Automated by .github/workflows/ci-benchmark.yml — regenerate locally with python -m tests.benchmark.run and neuralmind benchmark --quality.