v2.6.1b3

This release focuses on hardening the lyrics detection stage against AI formatting errors and improving the efficiency of the web search pipeline.

Prompt Engineering:

• Lyrics Detection (v4):
  • JSON Syntax Guard: Introduced a critical safety layer to prevent "field leakage," where the AI might accidentally include JSON keys or structural markers inside string values.
  • High-Efficiency Search: Re-engineered the execution pipeline to explicitly mandate simultaneous, multi-query web searches. This significantly reduces the number of AI turns required, lowering both latency and API costs.
  • Structured Output Format: Refined the prompt's output block to provide a clearer template for the model, ensuring consistent JSON generation.

Backend & Validation:

• Robust Timestamp Parsing:
  • Refactored _parse_timestamp_string_ms to use regular expressions. This allows the system to successfully extract valid timecodes even if the AI response contains "noisy" prefixes or suffixes (e.g., "01:23.456,start:") within the timestamp field.
• Data Model Enhancements:
  • Original Language Tracking: Added a dedicated original_language field to the Scene model to better capture and track the primary language of detected songs.
  • Scene Integrity: Implemented a Pydantic model_validator for the Scene class, ensuring that all detected scenes maintain chronological integrity (start time must be strictly before end time) before processing continues.

Bug Fixes:

• Fixed an issue where "field leakage" in AI responses could lead to JSON validation failures.

Full Changelog: 2.6.1b2...v2.6.1b3