fix(provider): never force a tool choice on models that refuse one

[filipeforattini]

Follow-up to #369 (ProviderTransform.supportsForcedToolChoice): audit of every place that forces a tool on a provider.

Changed

• packages/redcode/src/session/llm.ts: the one boundary every legacy session request crosses (AI SDK and native runtimes) now downgrades toolChoice: "required" to "auto" for models that refuse a forced tool choice, using the model's declared router parameters. prompt.ts already did this for structured output; this covers any other caller.
• packages/redcode/src/agent/agent.ts: Agent.generate used generateObject, which forces a JSON tool wherever the provider lacks native structured output (Vertex Anthropic sets supportsNativeStructuredOutput: false; gateway/combo ids without claude- fall back to the JSON tool). For models that refuse a forced tool choice it now uses prompted JSON (generateText) validated against the schema, with one repair retry.

Audited, already safe

• Semantic.generate/transform (core): tools: [], no tool choice.
• Title, summary, goal judge (legacy + core): plain text, no tool choice.
• Legacy compaction: toolChoice: "none" or undefined. Core runner: "none" on the last step only.
• Structured output in prompt.ts: already guarded by fix(provider): complete support for Claude Opus 5.5 and GPT-6 #369.
• packages/llm protocols (anthropic-messages, openai, bedrock, gemini): faithfully lower whatever tool choice the request carries; the guard lives at the call sites.
• LLM.generateObject in packages/llm forces a named tool but has no production callers (examples/tests only).

Tests

• test/session/llm.test.ts: Opus 5.5 with toolChoice: "required" sends tool_choice: {type: "auto"}; Opus 4.6 still sends {type: "any"}.
• test/agent/generate.test.ts: Opus 5.5 agent generation sends no tools/tool_choice, repairs a non-JSON reply once and decodes a fenced JSON reply.

Validated by CI only.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}