docs(validator): separate backend failures from probe calibration

[halfaipg]

What

Update the original failed-group audit with deployed token-limit v2, controlled stop/reasoning tests, direct DeepSeek and GPT-OSS tool-format comparisons, and an independent oracle for the four logic failures.

Evidence

• Read-only bounded PostgreSQL queries retain the original twelve larger-model tool failures and four larger-model logic failures.
• Six direct DeepSeek controls reproduce three malformed argument strings byte-for-byte; auto mode leaks native markup despite JSON arguments.
• Six actual GPT-OSS controls return five missing calls and one correct first-stage call. Three failures contain the correct argument object as text, not a structured call. A separate failed helper attempt with an empty auth header is retained and excluded.
• Four non-reasoning Qwen controls validate stop behavior on the owned backend, not fleet-wide compatibility.
• Independent arithmetic parsing verifies all four logic commitments and confirms wrong-integer answers.

Verification

Independent retained-response parsers and expected-commitment comparisons; offline deployed-worker request-builder check; diff-check and staged secret scans passed. Updated-head full CI remains required. Private prompts, responses and credentials stay outside git.

Risk and next steps

Docs only: no services, scores, historical records or money state changed. No permissive JSON scoring or blanket provider options. Endpoint-reported versions are not verified runtime artifacts. Actual backend host access, pinned backend qualification and full tool-chain tests remain necessary. Compensation budget, independent-operator reviews and human consent are still pending.