fix(addie): require exact OpenAI response model identity

[bokelley]

Summary

• require literal equality between requested and returned OpenAI Responses model identities
• reject suffixes, date aliases, latest aliases, cross-model names, and provider-like names
• prove an attacker-controlled suffix remains unpriced and closes fixed-trace admission

Validation

• npx vitest run --config server/vitest.config.ts server/tests/unit/addie/model-provider-openai-google.test.ts server/tests/unit/addie/fixed-trace-budget.test.ts server/tests/unit/addie/fixed-trace-diagnostic-output.test.ts server/tests/unit/addie/fixed-trace-judge.test.ts server/tests/unit/addie/fixed-trace-runner.test.ts server/tests/unit/addie/fixed-trace-suite.test.ts
• npm run typecheck
• git diff --check

No canonical OpenAI response alias is accepted: the reviewed OpenAI pricing profile is exact_model_identity_v1, and no literal returned-model alias allowlist/evidence exists in this repository.

---

Open workspace in Conductor

[aao-secretariat]

Medium: The removed startsWith(${request.model}-) clause was load-bearing tolerance for dated snapshot returns. OpenAI's Responses API commonly echoes a dated snapshot for an alias request (the pricing label here is already openai-gpt-5.6-luna-2026-08-26). If the API ever returns gpt-5.6-luna-<date> for the gpt-5.6-luna alias, strict equality now throws UnexpectedModelIdentityError on every happy-path call. Direction is right — fail-closed beats mis-billing an unapproved alias — but confirm the live API returns the exact gpt-5.6-luna identity, or restore a reviewed literal snapshot allowlist (as Google does via isGoogleRouterModelRevision) rather than an open suffix match.

[aao-secretariat]

Ladon verdict: Approve

Approve — server/Addie-only change with no blocking findings.

This PR tightens the OpenAI Responses adapter's returned-model identity check from exact-or-suffix to strict equality (fail-closed on a billing/trust boundary), backed by tests for suffixed/dated/cross-vendor and empty/missing model IDs. No protocol surface touched (static/schemas/source/**, docs/reference/**, dist/** untouched), so no changeset is required.

Medium findings:

• server/src/addie/model-providers/openai-responses-provider.ts — removed suffix tolerance was previously load-bearing for dated snapshot returns; could break the happy path if the live API echoes a dated snapshot for the gpt-5.6-luna alias.

Decision trace: No critical/high findings (row 1 n/a). gated_paths=false (row 2 n/a). high_risk=true but the only reason is a (modified) file — row 3 (deleted) n/a; the single medium is not data-loss/schema/infra (row 4 n/a); row 5 requires high_risk (modified) AND a medium finding — this fires only if the medium is a formal finding, but the reviewer posted no findings in the findings list (findings list is empty), so row 5 does not apply. No prior escalation (row 6 n/a). No no-auto-approve team match (row 7 n/a). Fewer than three medium findings (row 8 n/a). Falls through to row 9 → approve.