Add exec-harness cache assertion primitives #32

Objective

Add the harness primitives needed for deterministic cache-sensitive scenarios before implementing skills-cache-continuity.

Finish Line

Exec harness supports deterministic token/cache/prompt-prefix assertions needed by cache-sensitive scenarios.

Current Status

State: Active. Next action: inspect existing evaluate_expectations and fake Responses fixtures to design the smallest assertion extensions.

Scope

In:

• Allow fake Responses fixtures to supply usage values.
• Add expectation support for input token bounds, cached token minimums, and cache ratios.
• Add either prompt-prefix comparison support or an explicitly documented weaker substitute.
• Consider a no-fail-fast mode so scenario runs report all failures.

Out:

• No production Codex behavior changes.
• No old Every Code harness feature expansion unless needed by the cache scenario.

Acceptance Criteria

• Fake Responses SSE completion can carry per-response fixture usage.
• Expectations can assert cached input token minimums and cache ratio by turn.
• Expectations can catch catastrophic prompt bloat.
• Prompt-prefix stability has a deterministic assertion path.
• Unit tests cover the new harness assertion behavior.

Relationships

Child of #28. Blocks the skills-cache-continuity scenario child issue.

Validation

Run the exec-harness unit tests that cover fake Responses and expectation evaluation.

Decisions

Cache-sensitive behavior should be tested with deterministic fake usage first; speed-only signals are not enough.

Open Questions

What prefix length and cache-ratio thresholds are realistic across supported model/API behavior?