2.6.1b2

This release refines the AI prompting strategy to improve formatting reliability and better handle visual-only cues.

Prompt Engineering:

• Lyrics Detection (v3):
  • JSON Syntax Guards: Introduced strict rules for escaping, newlines, and key separation to ensure valid JSON output under all conditions.
  • Simplified Pipeline: Streamlined the metadata resolution steps for improved efficiency.
• Subtitle Generation (v15):
  • Dual-Trigger System: Clarified that subtitles are triggered by both vocal audio and prominent on-screen text.
  • Visual Events Exception: Explicitly instructs the AI to subtitle relevant visual text (like chapter titles) even when no audio is present.
  • Anti-Hallucination Hardening: Strengthened rules for ignoring incorrect or incomplete reference data and mandated manual transcription when the JSON reference fails.
  • Standardized Analysis: Updated examples to use a consistent, structured global_analysis format.
  • New Visual Text Example: Added a dedicated example for handling silent title cards.

Backend & Validation:

• Timestamp Validation:
  • Refactored timestamp parsing into an internal utility function _parse_timestamp_string_ms.
  • Implemented Pydantic model_validator on both Subtitles and Scene models to verify timestamp format and ensure the start time is strictly before the end time.
  • Improved robustness against malformed AI-generated strings (e.g., truncated JSON noise in timestamp fields).
• Static Analysis Fixes:
  • Resolved CliPositionalArg to Path type mismatch for static analysis.
  • Added explicit type casting in config.py and main.py to address errors where CliPositionalArg was not assignable to Path for methods like .resolve(), .stem, and .name in Pyright/mypy.

Full Changelog: v2.6.1b1...2.6.1b2