[Blaizzy]

A short message can fail with “Lower Max output to leave room for conversation compaction” when a saved Max output setting is equal to or larger than the effective context window. For example, context 4,096 and Max output 4,096 should allow “Hello” to generate a short answer.

Trigger compaction from actual input usage at the selected percentage (75% by default), without subtracting Max output or changing its saved value. Document fitting also stops reserving unused output tokens when Auto compact is enabled. The server counts and, when necessary, compacts the input before limiting generation to the smaller of Max output and the remaining context capacity.

Pin the bundled mlx-vlm server to the companion fix in Blaizzy/mlx-vlm#2458 (0a46548aab8fd034a68f54f5cb49cc6ce113728a). Preserve MLX_VLM_SOURCE_PATH overrides when resolving the pinned requirements. External servers need the companion fix as well. Compaction still happens between generation requests, including tool rounds; reaching the context limit ends a response with a length result.

Validation:

• Xcode app build and 36 focused compaction/document tests passed. Used a filtered test target because the repository's full test target has an existing SystemMonitorIdentityTests import/configuration failure.
• Bundled-server source override regression test passed; pip resolved/downloaded the exact VCS pin with the bundle's --only-binary=:all: option.
• Companion server suite: 541 passed, 6 opt-in integration tests skipped.
• Separate live MiniCPM5-2B HTTP checks with context and Max output both 4,096: streaming and non-streaming short prompts succeeded (19 input + 3 output); a 4,064-token prompt generated exactly 32 tokens and returned response.incomplete.

Addresses #670.