[zzwong]

Summary

Share report-identity preparation, sequential inference, receipts and private artifact publication between offline benchmarks and the separate diagnostic observer. System One request construction is shared; OpenAI Decisions and Clef retain their provider-specific adapters.

Direct and saved-candidate runs resolve keys through the configured credential store and provider-specific key mapping. Output reservation and complete-fixture size checks happen before credential lookup or provider calls. Retained inputs preserve complete report text and replayable request encoding.

Includes the credential and replay contract fixes. Post-review observer wiring remains in the separate shadow feature; this change does not alter live review verdicts or posting.

Validation

• make check passed, including tidy, formatting, lint, the full test suite and build.
• Hermetic command tests cover configured stores, direct and saved candidates, missing or conflicting credentials, complete request contents and zero-call rejection paths.
• Boundary regressions independently encode the expected fixture, accept the exact size limit and reject one byte over before credentials or inference. Provider accounting preserves unknown cache-write costs.
• Validation uses local fixtures and mock providers.

[zzwong]

The shared execution path preserves request and receipt behavior. Reservation failures stop before credential lookup or provider calls.