release: v0.10.1 — graceful per-frame zoom skip

[JRemitz]

Summary

A single OpenAI vision-call timeout on one frame previously aborted the entire smart-zoom analysis and failed the render. User reported this twice in a row (frame_0005.png timed out 3 retries in 30s each), losing both attempts at a render.

With the Catmull-Rom smoothing added in reeln-cli#30 (v0.0.40), the renderer interpolates across a missing keyframe with no visible discontinuity. Dropping one of eighteen frames is now a no-op.

Changes

• _analyze_frames_for_zoom catches OpenAIError per frame, logs the skip, and continues with the remaining frames.
• New config fields:
  • min_zoom_points_success_ratio (default 0.5) — minimum fraction of frames that must succeed.
  • min_zoom_points (default 2) — absolute minimum number of successful keypoints.
• Failure (with the existing error signal in context.shared["smart_zoom"]) is only raised when both thresholds aren't met.
• Strict pre-0.10.1 behaviour can be restored by setting min_zoom_points_success_ratio: 1.0.

Why these defaults

The renderer's spline needs at least 2 points to interpolate. 50% is a sensible floor — below that the path no longer reliably tracks the action. For typical zoom_frames=18 usage, that means up to 9 frames could fail and the render still completes; in practice timeouts hit ~1 frame, so the user gains a huge resilience improvement at zero correctness cost.

Test plan

• uv run pytest -q — 293 tests pass (4 new)
  • test_one_frame_failure_continues_analysis — exact regression: 18 frames, one fails → 17 keypoints in path, no error
  • test_failure_below_ratio_threshold_raises — 1/4 success → error
  • test_ratio_threshold_configurable — ratio=1.0 restores strict mode
  • test_all_frames_failing_signals_error — full failure still surfaces
• Manual: re-attempt the failing render — should now complete with one logged "skipping" warning instead of erroring out

🤖 Generated with Claude Code