feat(openai,dmr): switch thinking off on OpenAI-compatible endpoints and add provider_opts.extra_body

[IsmaelMartinez]

Open-weight reasoning models served through OpenAI-compatible endpoints think by default and each server exposes a request field to turn it off, but the openai client only honoured thinking_budget for OpenAI reasoning model names, applyModelDefaults dropped a disabled budget before any client saw it, and no provider_opts beyond the sampling allowlist reached the wire, so a Qwen3 model on mlx_lm, llama.cpp, vLLM or SGLang reasoned on every call and could spend the whole max_tokens budget before answering. This adds provider_opts.extra_body, an object merged verbatim into the chat-completions body for the openai and dmr clients on any provider (an explicit key wins over anything derived), and makes thinking_budget: none/0 on a user-supplied base_url send chat_template_kwargs.enable_thinking=false on Chat Completions with the existing 256-token output floor; the dmr client sends the same per request alongside llamacpp.reasoning-budget, which now actually reaches _configure for none. The switch is strictly opt-in: it needs the user's own none/0 (resolved once in the factory as options.WithChatTemplateThinkingOff, so title-generation and compaction clones inherit it but cannot trigger it alone), compares the resolved base_url against the alias default so built-in aliases and their clones are unaffected, and is never sent for azure, chatgpt or OpenAI model names behind a proxy. Docs, agent-schema.json and examples/thinking_budget.yaml are updated, including a note that reasoning tokens count against max_tokens.

Refs #4309

[lyonsno]

I reproduced a DMR side effect of preserving the disabled budget: MCP sampling clones now send llamacpp.reasoning-budget: 0 to the model-wide _configure endpoint.

Using the real registry and CloneWithOptions with the options from samplingModelOptionsFor, an agent configured with budget 4096 produces:

base 032d705: initial configure 4096; sampling configure omits budget
head 6e0a73b: initial configure 4096; sampling configure sets budget 0

Title/compaction clones skip that second configure call; sampling clones don't. Model Runner stores accepted configuration by model, so on the llama.cpp route this can leave the main agent using a zero reasoning budget after sampling. The original provider's config still says 4096, and its completion path doesn't reconfigure the runner.

Could sampling clones also avoid model-wide reconfiguration and use the new per-request thinking-off field? I verified the HTTP writes on both revisions; the downstream effect is traced through Model Runner's scheduler/loader, not a live-model run. The parent already reconfigures with the budget omitted—the newly introduced part is explicitly setting it to zero.

[aheritier]

👋 This PR has merge conflicts with the base branch. Please rebase or merge the latest base branch and resolve them. I've moved it to draft and added status/needs-rebase; it'll be picked back up automatically once the conflicts are cleared.

[IsmaelMartinez]

Applied in a38a765dc. Model Runner replaces a model's stored configuration on every _configure, so the parent's budget-omitted reconfigure already wiped the agent's 4096 on main (docker model configure show → Config: {}, verified live on llama.cpp b9879); this PR changed the wiped value to 0. dmr.NewClient now skips _configure for NoThinking clones, matching title and compaction clones, and the sampling request keeps its per-request chat_template_kwargs. WithNoThinking is set only by the sampling and title clones, so nothing else changes. TestNewClientClonesSkipConfigure covers the main agent (configures 4096) and the three clones (no configure); through the real registry the scenario now produces one configure call instead of two.