fix(responses,google): deliver structured output to routed chat models and honor effort ladders

[lidge-jun]

Summary

Routed openai-chat models now receive structured-output requests instead of silently returning prose, and any Google model with a declared effort ladder gets its thinkingLevel sent.

Structured output. parseRequest reduced the Responses text.format object to a boolean (_structuredOutput), so the schema, name, and strict flag were discarded at the parser and no adapter could forward them — src/adapters/openai-chat.ts had no response_format handling at all. The format now rides options.textFormat and is re-nested under json_schema on the Chat wire, the inverse of responseFormatToText. A caller that asks for a schema and gets prose back will try to JSON.parse it, which is the failure this removes.

Gemini effort. thinkingLevel was sent for a narrow model set only; it now follows any declared effort ladder, with the docs stating that both reasoningEfforts and modelReasoningEfforts assert capability and that CCA uses its envelope-specific path.

Attribution

Both fixes are @DevMello's (Pranav Yerramaneni) work, from #985 and #978. Seven of the nine commits here are cherry-picked with their authorship intact; the two I authored carry Co-authored-by: Pranav Yerramaneni.

One correction on top of #985: its json_schema branch was guarded by textFormat.schema !== undefined, so a schema-less json_schema fell through and no response_format was sent — silently reproducing the original defect on a narrower input. The guard now sits on the schema member instead of the branch, with a regression test asserting response_format.type === "json_schema" is still sent when no schema is supplied.

#985 and #978 are left open for the author. Planning unit: devlog/_plan/260806_stacked_bug_campaign/ (phases 090, 100).

Stack 5 of the 260806 attribution campaign, stacked on #1136.

Verification

• bun test tests/openai-chat-hardening.test.ts tests/responses-parser.test.ts tests/chat-completions-endpoint.test.ts tests/google-hardening.test.ts — 129 pass, 0 fail
• bun run typecheck — exit 0
• bun run privacy:scan — passed
• bun run test (full) — exit 0
• Docs build — 221 pages built
• Full suite also passed the repository prepush gate on push

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults.

[github-actions]

✅ Deterministic PR hygiene checks passed.