Fix native OMP eval catalog compatibility and tool evidence

[0xashrk]

Summary

Correct native OMP live-eval compatibility and accounting without changing the suite, prompts or grader. No new OAuth Responses shim is introduced; the existing native OAuth path is used.

• Accept the exact canonical 31-tool research catalog or exact connected 32-tool catalog, still rejecting missing, unknown or duplicate names. Expose/record only the 31 research tools; exclude gina.renderReadOnlyDashboard.
• Permit legitimate JSON Schema union types in AJV while retaining strict validation and no evaluator-side argument coercion.
• Explicitly launch OMP with --tools read; the ACP builtin declaration alone does not enforce runtime tool filtering. Supply a required string path read schema rather than depending on a fragile title.
• Keep native read activation/failure evidence separate from canonical research routing. Reads are not confined to staged skills or a filesystem sandbox.
• Correlate native MCP mirrors to captured host executions one-to-one using trusted server/tool/correlation metadata, equal arguments and matching success/error outcome. Retain every real execution, surplus/mismatched mirrors and unknown native calls; do not blanket-drop acp_* calls.

Verification

• Six focused regression tests: five fail against the previous source; all six pass with the fix. Coverage includes real repeated executions, surplus mirrors, argument mismatch, unknown native evidence and failed native reads.
• Evaluation package: 236 tests passed. Root typecheck and build passed.
• Actual OAuth metadata preflight verified false duplicate routing before the fix and one canonical execution afterward.
• Two real production-MCP spot runs, four cases × three repetitions, clean chats, medium reasoning, native OpenAI OAuth (no OpenAI API key), separate Gina OAuth tools:read grant. Same OMP 18.1.17 executable, harness-acp 1.0.40, suite and grader.

Recorded metric	GPT-5.5 / medium	GPT-5.6 Sol / medium
Implemented-grader overall pass	11/12	11/12
Routing	11/12	11/12
Arguments / completion	12/12 each	12/12 each
P50 latency	26.633s	27.949s
P95 / max latency	54.557s	55.121s
Reported total tokens	185,084	335,030
Within declared 30s target (diagnostic)	8/12	6/12
Max reported result bytes	76,565	76,565

Both models failed one AAVE metadata trial for two actual spot.getTokenMetadata executions: GPT-5.5 repetition 3 and Sol repetition 1. Neither had a recorded tool-call error. Those failures were retained, not replaced; the stored evidence does not explain the extra invocation.

Grading limits — not broader quality or safety claims

• Only the three simple-price trials assert specific argument values (ids=ethereum, vs_currencies=usd); the other nine argument scores automatically pass for lack of a case-specific constraint. Host schema validation is separate, and OMP may coerce arguments before forwarding.
• Completion checks completed status and recorded tool errors, not answer semantics. Final-answer grounding, chart presentation and correctness are not graded.
• The YAML's 30s / 100,000-byte budgets are not pass/fail gates.
• Safety and skill activation are N/A/ungraded, not passes. Observed forbidden tools/scopes still fail safety, but missing requested_scope evidence can omit its score.
• Native reads are allowed; this is not filesystem confinement, Codex CLI plugin-install verification or an MCP App rendering test.
• This small live sample is not a general model ranking or monetary-cost comparison. Usage totals exclude setup/debug work. Full answers and raw tool argument/result bodies were not retained.

Reproducibility / evidence

• GPT-5.5 run: spot-openai-oauth-20260911T135127Z; aggregate SHA-256 d12167198c14054d5b0599c9fd7fc3758a2d9c1aaa784009ae7352fb91fdfb0f.
• Sol run: spot-openai-oauth-sol-20260911T142646Z; aggregate SHA-256 4f4d0f5a544905aa200abebdfec431bc93daba21ce95d3710f2d3a82d6a64a71.
• Tested source: 5f98d54cf0f6514c90116a4c7c1d889cdd7dc485; Sol tip 780dc809e9956a329d9e60c72503de449477699b changes only one README sentence.
• A three-page comparison PDF and checksummed archive were delivered separately to the requester: both original reports, all 24 valid attempts, journals, normalized tool-name/status traces, exact suite/grading snapshots and source patch. They are not committed or publicly attached to this PR. Temporary credential profiles/files and diagnostic wrapper were removed; no credentials are included.

These changes repair runner evidence, not the grading-coverage limitations listed above.