fix(llm-client): disable reasoning per-model via baseURL-aware table

[DerekEXS]

Extends the single-line /qwen3|deepseek.*r1|qwq/ regex (issue #929) into a baseURL-aware rule table that injects the correct disable-thinking request parameter for each known Chinese reasoning-model family on its respective provider endpoint.

Why a table (not just a regex fix)

The same model can require different parameter shapes on different providers. For example, deepseek-v4-flash uses chat_template_kwargs.enable_thinking=false on Aliyun DashScope but thinking.type="disabled" on DeepSeek's own API. Model name alone is not enough — the table selects the right parameter per (model family, provider).

Coverage (parameters from each provider's official docs)

Family	Provider	Parameter
Qwen3 / QwQ	DashScope / vLLM	chat_template_kwargs.enable_thinking=false
DeepSeek V4 / R1	DashScope	chat_template_kwargs.enable_thinking=false
DeepSeek V4	DeepSeek official	thinking.type="disabled"
GLM-4.6 / GLM-Z1	Zhipu BigModel	thinking.type="disabled"
Doubao thinking/seed	Volcano Ark (OpenAI-compat)	thinking.type="disabled"
Kimi K1.5 / K2	Moonshot	thinking.type="disabled"
MiniMax M2 / abab	MiniMax	thinking_budget=0
Hunyuan T1 / Turbo-S	Tencent	enable_thinking=false
StepFun step-3 / step-2	StepFun	thinking=false

API

• New buildDisableReasoningParams(model, baseURL) returns the correct params or null.
• shouldDisableReasoningForJson(model) is kept as a backward-compatible boolean wrapper (used by existing tests).
• recoverJsonFromReasoning (PR fix(smart-extractor): only learn noise from a genuine empty extraction; recover JSON from the reasoning field #914) remains as a safety net for models/providers not yet in the table.

Verification

Live test against DashScope deepseek-v4-flash with the extraction prompt, three variants:

Variant	Reasoning chars	Completion tokens	Total
(baseline) no param	225	67 (incl. 53 reasoning)	102
chat_template_kwargs.enable_thinking=false	0	12	47
thinking.type="disabled"	0	12	47

Completion tokens -82%. DashScope deepseek-v4-flash accepts both parameters. Only this one model lane was live-verified; the rest are doc-sourced and rely on recoverJsonFromReasoning as a safety net for unverified models/providers.

Tests

test/llm-api-key-client.test.mjs — all 8 pass (6 original + 2 new: per-provider param table, deepseek-v4-flash end-to-end injection).

Fixes #929

[app3apps]

The DashScope deepseek-v4-flash fix is valuable, and the verified path looks sound. I am requesting changes because the expanded table also ships provider/model mappings that are incorrect for configurations this project supports.

1. MiniMax M2 cannot be disabled by this rule (src/llm-client.ts:188-192). The current MiniMax OpenAI-compatible documentation explicitly says thinking cannot be disabled for M2.x, but this regex matches every MiniMax-M2* model and injects thinking_budget: 0. That cannot deliver the advertised behavior and may be ignored or rejected. Please remove/narrow the M2 and broad abab mapping unless there is a provider-specific documented endpoint and request-level verification: https://platform.minimaxi.com/docs/api-reference/text-openai-api

2. Most entries are not actually provider-aware (src/llm-client.ts:164-207). Seven rules use a universal default, so private provider fields are sent to every OpenAI-compatible endpoint after a model-name match. There is a concrete mismatch with the built-in SiliconFlow path: SiliconFlow documents top-level enable_thinking: false for zai-org/GLM-4.6, while this table sends { thinking: { type: "disabled" } }. At best reasoning remains enabled; a strict gateway may reject the request before recoverJsonFromReasoning can help. Please positively identify provider hosts and return null for unknown endpoints, or add an explicit capability override. Add negative unknown-provider coverage and a SiliconFlow case: https://docs.siliconflow.cn/en/api-reference/chat-completions/chat-completions

3. The Kimi matcher includes its own unsupported exception (src/llm-client.ts:179-185). The comment says kimi-k2.7-code cannot disable thinking, but /kimi.*k[12]|kimi.*thinking/i matches it and injects thinking.type=disabled. Please use a disable-capable allowlist or add an explicit exclusion plus a negative test.

A clean way forward would be to narrow this PR to the live-verified DashScope DeepSeek V4 fix, then add the other families only with provider-specific evidence and tests.

Verification performed: full npm test passed on the PR branch; npm run build passed with no generated diff; the merge result with current master also built and passed the combined LLM API-key client tests.