fix(llm): retry a sub-call without response_format when the endpoint refuses structured outputs

[plombeer31]

What

Before: cloud memory sub-calls (query rewriter, link generator, vote runner, consolidation distill) always sent OpenAI Structured Outputs (response_format: {type: "json_schema"}). An endpoint with no structured-output support refused every one of them — OpenRouter pinned to Z.AI's own endpoint for z-ai/glm-5.3-flash answers 404 No endpoints found after its "Filter by Parameters" routing step; direct vendor APIs without json_schema answer 400. The sub-call failed forever on that provider, and because every cloud OpenAiHttpError classifies transport, each refusal also ran through the provider fallback chain.

After: when a request carrying responseFormat is refused as a structured-output refusal, OpenAiProvider.complete retries it once on the same provider without response_format, returns that result, and remembers the (provider id, model) pair for the rest of the process, so later sub-calls to it skip response_format up front. One warn line the first time. A handled refusal never reaches runWithFallback, so it never advances the chain or trips its breaker. Streaming main-turn requests never carry responseFormat and are untouched.

Why

The sub-call parsers never needed Structured Outputs to get an answer: all four (rewriter, link, vote, distill) try JSON first and fall back to the text format their prompt asks for. Losing decode-time enforcement degrades an answer; a permanent 404 loses it.

The detector is deliberately narrow — an OpenAiHttpError that is not our own timeout, and:

• 404 whose body says No endpoints found together with requested parameters, a feature word (response_format / json_schema / json_object / structured outputs), or a "Filter by Parameters" funnel step that actually dropped endpoints (the step appears in every OpenRouter funnel, so its count is checked when readable);
• 400 / 422 whose body names a feature word — except Alibaba's must contain the word 'json' (a prompt problem on an endpoint that does support structured outputs; fix(memory): structured-output sub-calls that OpenAI strict mode and Qwen reject #414 fixes the prompt) and request-size rejections.

401 / 402 / 403 / 429 / 5xx, network failures and untyped errors are never refusals. The pair is remembered only after the retry succeeds; a retry that fails propagates its own error and remembers nothing.

How it was verified

• npm run lint clean
• npx vitest run src/llm/provider/openai src/runtime/llm-fallback-seam-structured-output.test.ts src/runtime/llm-fallback-seam.test.ts src/llm/fallback — 26 files / 435 tests green
• new tests: a 404 refusal sends a second request identical except for response_format and returns its success; at the seam advanceFrom is never called and the local link is never reached; the next call skips response_format with no failed round trip; memory is per provider id and per model; invalid-key, context-length and json-word 400s propagate after a single request; no retry without responseFormat or with tools; a failed retry remembers nothing; completeStream never retries; a control test shows the same 404 without responseFormat still advances the chain
• vacuity: forcing the detector to return false turns 11 of the 34 new tests red across all three files; removing the must contain the word json exclusion turns 2 red (that 400 must not be read as a refusal)
• live, this branch's build, z-ai/glm-5.3-flash pinned to Z.AI's endpoint, same scripted 8-turn session as the baseline:

	v0.6.0	this branch
rewriter outcomes	failed 18	ok 11, timeout 7, failed 0
vote runner	8 fired / 8 failed	8 fired / 8 ok, votes applied
"No endpoints found" log lines	26	0
structured-output warn lines	–	1
provider switches	–	0

Merge note: #413 also edits OpenAiProvider.complete (it threads providerPreferences into the buildOpenAiChatBody call this PR moves into a closure) — whichever lands second needs a trivial rebase. #414's JSON instruction is only added when response_format is attached, so the stripped retry sends the prompt unchanged.

The 7 remaining rewriter timeouts are the 3 s rewriter budget on a thinking model — before, the call failed in ~74 ms; now it actually waits for GLM. #416 raises that budget to 10 s and runs the rewriter once per turn.