fix(llm): prompted JSON for models refusing forced tools, mid-stream quota classification

[filipeforattini]

Plan wave 1, item 3 (llm-forced-tool-and-quota): A4 + A5.

A4: LLM.generateObject without a forced tool choice

• New pure LLM.supportsForcedToolChoice(modelID, declared?) in packages/llm/src/llm.ts (Claude Opus >= 5.5, Fable, Mythos refuse; declared === false wins). ProviderTransform.supportsForcedToolChoice in redcode now delegates to it (redcode -> llm dependency direction).
• ModelCompatibility.forcedToolChoice lets a model or router declare the refusal on an llm Model.
• For refusing models, generateObject sends no tools/tool_choice, appends a system instruction with the JSON Schema, decodes the reply (fences stripped) against the same schema, and repairs once (same shape as fix(provider): never force a tool choice on models that refuse one #377 for Agent.generate). A second failure is an InvalidProviderOutput LLMError.

A5: mid-stream quota / content-policy classification

• New shared streamProviderError in packages/llm/src/provider-error.ts reuses isQuotaFailure / isContentPolicyFailure (the fix(provider): classify gateway quota exhaustion and 402 as non-retryable #378 classifiers). A stream has no status, so throttle shapes (rate_limit_error, rate_limit_exceeded, Bedrock throttlingException) stand in for 429.
• Anthropic Messages, OpenAI Responses and Bedrock Converse stream parsers route their provider-error events through it: quota -> classification: "quota", retryable: false, message with the shared QUOTA_HINT; content policy -> classification: "content-policy", retryable: false.
• ProviderFailureClassification gains quota and content-policy.
• QUOTA_HINT moved to llm and reused by provider/error.ts.
• Legacy processor: a refusal-classified provider-error becomes a non-retryable ResponseStreamError (refusal), which MessageV2.fromError records as metadata.classification, and SessionRetry.retryable never retries it, whatever the message pattern says. (Before, it became NamedError.Unknown, and the quota hint's "exhausted" would have read as an overload.) The v2 runner already never retries provider-error events.

Tests

• packages/llm/test/generate-object.test.ts: capability table, prompted path, one repair, failure after repair, declared refusal.
• Per protocol: anthropic-messages, openai-responses, bedrock-converse quota/content-policy events.
• packages/redcode/test/session/retry.test.ts: classified stream refusal is not retried.

---

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