fix(anthropic): map model_context_window_exceeded to a length finish

[erzaehlsalex]

What

• mapStopReason maps model_context_window_exceeded to finishReason: 'length'. It was 'unknown'.
• Upgrades @anthropic-ai/sdk from ^0.110.0 to ^0.114.0. 0.114.0 is the first version whose StopReason type includes the value, so the mapping needs no hand-written union or beta type.

Why

On Claude 4.5 and newer, the API accepts requests where input plus max_tokens exceeds the context window. If generation reaches the window limit, it stops with model_context_window_exceeded. Anthropic's docs say to treat that response as truncated.

#211 makes this stop reason routine: an omitted maxTokens becomes the model ceiling, so a long prompt on a 200K model (Haiku 4.5, Sonnet 4.5, Opus 4.5) no longer ends with max_tokens. With 'unknown':

• callers checking finishReason === 'length' miss the truncation
• generateObject() skips its truncation error and throws StructuredOutputParseError on the cut-off JSON

The fix is independent of #211 and applies to main as is. It was split out of #211 so that PR can release first.

Notes

• package-lock.json also picks up the workspace package versions (0.25.0 → 0.28.1), which were stale in the lockfile.
• Not reproduced against the live API. The behaviour rests on Anthropic's context-windows and stop-reasons docs.

Test plan

• npm run release:check
• npm run test
• New unit test: a stream ending with model_context_window_exceeded finishes with 'length'

🤖 Generated with Claude Code