fix(shadow): evaluator self-writes results/shadow_live.json (FIX-2)

[neuron7xLab]

Creator

Make scripts/evaluate_cross_asset_kuramoto_shadow.py self-write the full {"eval": ...} payload to results/shadow_live.json on every invocation, so downstream consumers (verdict scripts, trajectory loggers, external pipelines) read live state instead of silently parsing stale JSON.

Critic

• Edge-case 1: write happens AFTER all metric computation; if eval errors out earlier the JSON is not produced. Acceptable: this matches stdout behaviour — no partial-state JSON. Downstream code must check file presence + freshness.
• Edge-case 2: filesystem mtime resolution on coarse-grained FS (some NFS / FAT32) can collapse two sub-second writes to identical mtime. Mitigation: test sleeps 1.1s between runs.
• Edge-case 3: file is created with default permissions; if a strict umask is in force, downstream readers under different uid may fail. Acceptable: same risk profile as LIVE_STATE.json already on disk.
• Risk not covered: if a future refactor adds an early-return that bypasses the new write, the test still requires paper_state/equity.csv (skips on environments without it). CI runners without the spike skip the contract test — that's the documented limitation.

Auditor

Falsification gate (FIX-2): consecutive evaluator runs MUST produce results/shadow_live.json with strictly advancing mtime.

False-positive avoidance:

• The test asserts mtime_2 > mtime_1, not >=. Two writes with identical mtime → FAIL.
• The schema check ("eval" key present + 6 required sub-keys) prevents a degenerate fix where the file is touched but content not written.
• A hypothetical "atomic-write-via-rename" implementation that creates a fresh inode each time is also caught: Path.stat().st_mtime resolves the new inode.

Verifier

```
$ rm -f results/shadow_live.json
$ python scripts/evaluate_cross_asset_kuramoto_shadow.py >/dev/null
$ T1=$(stat -c %Y results/shadow_live.json)
$ sleep 1.2
$ python scripts/evaluate_cross_asset_kuramoto_shadow.py >/dev/null
$ T2=$(stat -c %Y results/shadow_live.json)
$ echo "T1=$T1 T2=$T2 advanced=$([ $T2 -gt $T1 ] && echo YES || echo NO)"
T1=1778065314 T2=1778065316 advanced=YES

$ python -m pytest tests/scripts/test_shadow_eval_self_writes_json.py -x -q
. [100%]
1 passed
```

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: aa2daf5c37

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Remove skip gate that disables contract test coverage

This skipif makes the new contract test no-op on environments without ~/spikes/.../paper_state/equity.csv, which is common in CI, so regressions in writing results/shadow_live.json can pass undetected. The evaluator already tolerates a missing paper-state ledger and exits successfully with empty live metrics, so gating the test on that file is stricter than the runtime behavior and defeats the purpose of this regression test.

Useful? React with 👍 / 👎.