v2.2.0 — Echo Bleed Reduction

Added

• Widened time window: Echo detection window expanded from 15s to 30s across all three detection layers. Catches delayed echo up to 30s apart. Recent segment retention increased from 60s to 90s.
• Audio envelope correlation in live pipeline: is_echo() now runs as tier 3 in the live transcription pipeline (after energy gate, before Whisper). Uses envelope correlation with mic_rms guard to catch echo that Whisper transcribes as different words per channel.
• Prefix/suffix extraction: When clause splitting can't isolate genuine speech (no punctuation delimiter), scans 1-4 word prefixes/suffixes. Recovers acknowledgments like "Okay", "Sure", "Yes" glued to echo without punctuation.
• SequenceMatcher for medium segments: Character-level sequence matching for 6-15 word segments catches echo where Whisper wording differences reduce word overlap below threshold. Routes through clause recovery before full removal.
• 20 new tests (261 total)

Changed

• deduplicate_segments() default time_window: 15s → 30s
• _is_text_echo() default time_window: 15s → 30s
• _retract_echo_you() default time_window: 15s → 30s
• Recent segment retention cutoff: 60s → 90s
• _extract_unique_clauses() now falls back to prefix/suffix extraction when clause splitting fails