fix(agent): support Sonnet 5 output limits

[iamjr15]

Why

Claude Sonnet 5 was added to the product catalog while @ai-sdk/anthropic remained on 3.0.84, which does not recognize that model ID. The provider therefore applied its unknown-model 4,096-token compatibility limit. Sonnet 5 defaults to adaptive thinking, so long structured tool calls could be truncated before required fields such as file content were emitted.

Decision

Upgrade only @ai-sdk/anthropic within its existing 3.x major to 3.0.107. This version recognizes claude-sonnet-5, uses its documented 128K output capability, forwards current thinking controls, and remains compatible with the existing AI SDK 6 and Mastra 1 stack. No retry, prompt, schema, or runtime workaround is introduced.

Architecture and migration effects

• No architecture boundary changes
• No database or migration changes
• Lockfile-only transitive updates under the Anthropic provider

Verification

• pnpm lint
• pnpm typecheck
• pnpm turbo build --force
• pnpm deadcode
• pnpm architecture:check
• pnpm turbo skills:build
• Focused typechecks for @cheatcode/agent-core and @cheatcode/agent-worker
• Captured the generated Sonnet 5 request with a mock transport and verified model: claude-sonnet-5 and max_tokens: 128000

Production flow QA will follow deployment.