Test whether detected MoE shifts justify cache action

[Labeeb2339]

Summary

• preregister a causal shift-actionability pilot before inspecting held-out seeds
• add frozen execution plans, shadow decisions, finite-oracle diagnostics, cluster bootstrap intervals, and provenance hashes
• publish the complete failed carry-forward result and surface it honestly in the workbench
• add regression and replay tests for the checked-in evidence

Frozen result

• controlled shifts detected: 30 / 30
• gated actions: 76
• harmful actions: 16 / 76 (21.05%)
• worst regression: +21.34%
• median benefit in the useful 64-slot subset: -9.75% modeled traffic
• oracle-actionable coverage: 1 capacity, below the required 3
• carry-forward decision: false

The detector passed, but the action failed its preregistered safety and coverage gates. This PR does not claim a new algorithm, quantization method, hardware measurement, or breakthrough.

Reproducibility

• protocol commit: daf79356365b21249fde42012f1b446069809548
• execution commit: 090481f4b710996a343e3b07332d69ff16c03258
• result SHA-256: be80f2238f8dd33587f0bf1f00656d3a963138368c82ca7f8840e9863cb4f681
• canonical payload SHA-256: 7b0b123d81344eb5b12e22f6539a266e00706b324e45160ff15854c3eca4f6c6
• independent byte-for-byte verifier passed

Validation

• npm run check — 42 unit tests, production build, and 2 rendered HTML tests passed
• npm audit --omit=dev --audit-level=high — 0 vulnerabilities
• desktop and 390 px mobile visual QA — no horizontal overflow or console warnings

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 34d13120a7

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]

Regenerate provenance for the submitted commit

In this reviewed commit, git log -1 --format=%H -- lib/simulator.ts lib/actionability.ts scripts/run-actionability-pilot.ts ... resolves to ef640752ef58e4b3f57f38863f95adeb1e14ae86, but the checked-in payload records 090481f4... as the implementation commit. Because --verify rebuilds provenance with latestCommitFor(implementationPaths) before doing the byte-for-byte JSON comparison, running npm run benchmark:actionability:verify from this commit will recompute ef640... here and reject the committed evidence; the same squash-history mismatch applies to the recorded protocol commit below. Regenerate the evidence/provenance on the commit being submitted, or keep the preregistration/execution commits in the submitted history.

Useful? React with 👍 / 👎.