fix(strix): classify provider tool-contract failures without masking findings

[seonghobae]

Incident

The exact-head Strix required check for .github PR #791 failed in workflow run 31071120319, job 92519148857, before producing a review report.

The controlled terminal signature was:

agents.exceptions.ModelBehaviorError: Tool execute not found in agent strix

The same log reported Vulnerabilities 0, no report artifact, and no source-level finding. This is a model/tool-contract failure in the read-only reviewer plane, not evidence that PR #791 is unsafe and not a condition that may be converted into a passing finding review without an explicit bounded policy.

Required test-first repair

Add permanent RED regressions before production changes that prove all of the following:

1. the exact controlled ModelBehaviorError / missing declared tool signature is classified as a distinct finite-cardinality provider-tool-contract outcome;
2. classification is permitted only when there is no generated vulnerability report, no finding record, and the scanner's controlled vulnerability count is zero;
3. any finding, partial finding artifact, configuration error, source checkout error, policy violation, malformed evidence, credential failure, or unknown exception remains fail-closed;
4. arbitrary model text cannot spoof the controlled signature;
5. bounded diagnostics include the stage and controlled class only, never prompt text, page content, credentials, headers, arbitrary exception text, or model output;
6. exact-head required-check policy remains explicit: either retry through an independently reviewed provider/model route, or publish a machine-readable review_unavailable result that cannot be treated as approval;
7. no product PR source is mutated merely to retrigger or conceal reviewer infrastructure failure;
8. existing reviewer credentials and identity separation remain unchanged.

Provider and credential boundary

• Do not introduce COPILOT_GITHUB_TOKEN as model authentication.
• Do not change the established read-only review-agent credential chain in this repair unless a separately reviewed migration is required.
• Any scheduled or write-capable repair worker must use OpenCode with NVIDIA_NIM_API_KEY; this issue concerns the independent read-only Strix review plane.
• Never synthesize APPROVED, merge, update branches, release, or weaken branch protection from an unavailable-review classification.

Verification

Require exact-head focused tests, the complete central suite, 100% production statement and branch coverage, 100% public docstrings, shell/Python compilation, Security Scan, Semgrep, CodeQL, secret scanning, and a current-head independent review before integration.

Operator evidence

Record the incident, classification table, retry/escalation path, rollback, non-claims, and APA 7th primary-source references in docs/doctoring/ and CHANGELOG.md.