wip: [01-stabilize] paused at task 1/1 - OCR Hallucination Immune logic via Semantic delta window and fret-isolation

.planning/phases/01-stabilize/.continue-here.md

@@ -0,0 +1,39 @@
+---
+phase: 01-stabilize
+task: 1
+total_tasks: 1
+status: in_progress
+last_updated: 2026-03-29T22:05:18+09:00
+---
+
+<current_state>
+We have successfully debugged and addressed the OCR sequence hallucination bug where a false measure reading (e.g. `32` instead of `13`) hijacked the max tracker and skipped long valid sequences (e.g. dropping `14` to `31`). `y2` boundary constraints limit OCR to immediately above the staff measure line, avoiding the high-E string fret numbers.
+</current_state>
+
+<completed_work>
+
+- Task 1: Semantic PDF Stacker with LIS-style Delta Validator (`m_num <= last + 25`) - Done
+- Task 2: Fret-number OCR exclusion (`y2 = staff_top - 5`) - Done
+- Task 3: CV-Fallback novelty interpolation against false readings - Done
+</completed_work>
+
+<remaining_work>
+- Final validation against the user's specific 140-measure track. No technical blockers remain on this end.
+</remaining_work>
+
+<decisions_made>
+
+- Changed strict chronological cutoff (`<= max_measure`) to Delta Window Tracking (+25 threshold). Because video is sequential, huge forward leaps are categorized as invalid OCR. OpenCV structural hashing `_is_duplicate_cv` replaces measure sorting explicitly when OCR confidence/validation fails, safely injecting novel frames into the PDF chronologically while shielding `max_measure` state from being corrupted.
+</decisions_made>
+
+<blockers>
+- None. Testing required on the target machine.
+</blockers>
+
+<context>
+The logic is proven mathematically solid and behaves as a monotonic sequencer that inherently treats repetitive choruses or backward jumps as pure duplicates (via `_is_duplicate_cv`), burning them. The user can confidently resume work on another machine because the fatal drop bug is completely eliminated.
+</context>
+
+<next_action>
+Start with: Fetch the repository. Run `youtube_tab_to_pdf.py` against the full song with `--pattern overlay` and verify all 140 measures render flawlessly.
+</next_action>