fix(openrouter): stop silently ignoring thinking effort off

[filipkujawa]

Summary

GOOSE_THINKING_EFFORT=off was a silent no-op on the OpenRouter path: apply_reasoning_config returned early and sent no reasoning config, so models that reason by default (Kimi K3, DeepSeek-R, GLM) kept thinking at full default effort.
On Terminal-Bench 2.1 that unbounded thinking cost goose 0/6 on cobol-modernization and build-pov-ray; disabling reasoning at the wire level flips them to PASS and halves median output tokens.

Now off on reasoning-capable models sends OpenRouter's documented disable knob {"reasoning": {"enabled": false}}, keeping the lowest-supported-effort clamp the upstream builder already computed for OpenAI-shaped models.
Some endpoints have mandatory reasoning (gpt-5 family, gemini-3.5, gpt-oss, deepseek-r1, some grok) and deterministically reject every disable spelling with a clear 400; rather than maintaining a name list that goes stale, the provider downgrades the request to {"reasoning": {"effort": "low"}} and retries once when the router reports that error.
The 400 is marked permanent so the generic retry loop does not burn attempts on it first.

Unchanged: an explicit reasoning object from OPENROUTER_PARAMETERS still wins (and never triggers the downgrade), and non-reasoning models still get no reasoning field.

Testing

• Unit tests assert the exact JSON per case: kimi-k3 off (resolved through the canonical registry), the kept gpt-5 clamp, non-reasoning models (including OpenAI-shaped chat variants like gpt-5.1-chat), and OPENROUTER_PARAMETERS precedence under off; a wiremock test drives the provider through the mandatory-reasoning 400 and asserts the downgraded retry.
• Live wire checks: {"enabled": false} yields 0 reasoning tokens on kimi-k3, glm-4.6, gemini-2.5-flash, grok-4.20, grok-4.3, claude-sonnet-4.5; mandatory endpoints (gpt-5-mini, gpt-5-pro, gemini-3.5-flash, gpt-oss-120b, grok-4.5, deepseek-r1) reject every disable spelling and all accept effort:"low".
• End-to-end with the patched binary on gemini-3.5-flash: request log shows the disable, the router 400, and the successful effort:"low" retry.
• Terminal-Bench 2.1 (patched binary, off, K3): the three signature tasks went 9/9 at k=3 vs 0/9 in thinking-on baselines; a 7-task no-regression slice all passed.
• cargo fmt, cargo clippy -D warnings, cargo test -p goose green.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 190763d756

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".