feat(llm-sdk): Anthropic forced-tool fallback for models without output_config

[Sahil5963]

Context

PR #95 added native output_config.format support for Anthropic structured output (Claude 3.5 / 3.7 / 4 series). This works on:

• Anthropic Claude API ✅
• AWS Bedrock ✅

It does NOT work on:

• Google Vertex AI (Anthropic models hosted via Vertex don't expose output_config)
• Older Claude 3 base models (claude-3-opus-20240229, claude-3-sonnet-20240229, claude-3-haiku-20240307) — capability flag set to false in PR fix(llm-sdk): structured output across providers + max_completion_tokens for o-series/gpt-5 (closes #93) #95

Problem

Customers on Vertex AI or pinned to older Claude 3 models will silently get plain text instead of structured JSON when they pass responseFormat. The capability gate warns but doesn't actually return JSON.

Proposed fix

Implement the forced-tool pattern (matches Vercel AI SDK's behavior):

1. When responseFormat.type === "json_schema" AND model lacks native output_config support, inject a synthetic tool whose input_schema is the user's response schema:

   tools: [{ name: "json", description: "Respond as JSON.", input_schema: <user schema> }]
   tool_choice: { type: "tool", name: "json" }

2. In response parsing, extract the tool call's input and surface it as content so the SDK behaves identically across all paths.

Reference

Vercel's implementation: packages/anthropic/src/anthropic-language-model.ts:336-352 (in their repo). They gate via structuredOutputMode: 'auto' | 'outputFormat' | 'tool'.

Files to touch

• packages/llm-sdk/src/adapters/anthropic.ts — add fallback path in buildRequestOptions + tool extraction in complete() / stream parser
• packages/llm-sdk/src/providers/anthropic/provider.ts — same in doGenerate / doStream
• packages/llm-sdk/src/providers/anthropic/provider.ts — flip older Claude 3 models from jsonMode: false → true once fallback is wired