[TheGreatAxios]

Closes #33

Stacked on #32 (base feat/account-id-quirk).

What changed

• InferenceOptions.effort becomes the Codex effort option before the request is built: low, medium, high as given; max as the model's highest listed level up to max (max, or xhigh on gpt-5.5, from openai/codex@b1e72963c3 supported_reasoning_levels), else high. gpt-6.1 is in CODEX_DEFAULT_MODELS but not in that catalog, so it falls back to high.
• A non-empty providerOptions[CODEX_REASONING_EFFORT_OPTION] still wins.
• The mapping is done here rather than through feat: map interchange's neutral effort to reasoning.effort corbits-openai-responses#33 so it works on the released @corbits/openai-responses 0.2.2; once Map Interchange's neutral effort to Codex reasoning.effort #33 ships, this can move to its reasoning.maxEffort quirk.
• README and AGENTS.md document the mapping.

Verification

• bun run check (typecheck, lint, format, 21 unit tests) passes.
• New tests pin the per-model max, the unlisted-model fallback, and option precedence (non-empty wins, none suppresses, empty falls through).
• Independent review findings addressed (empty-option precedence, comment accuracy, return type, tests, docs).

[corbits-triage]

Triage: In progress

Triage is waiting for the decision model's answer before it decides. This comment shows the verdict once it has.