feat: Phase 2 close-out — threshold alignment, DEVLOG, docs update

[AD2000X]

src/:

• layout_detector.py, table_detection.py: default threshold aligned to 0.3
• layout_parsing.py: DEFAULT_TABLE_SCORE/DEDUP_IOU named constants; shared defaults
• config.py: minor cleanup
• tatr_postprocess.py: empty row/col grid -> invalid; dedup_row_col_bands as default

scripts/:

• run_layout_batch.py, eval_layout_iou.py: read shared defaults from layout_parsing
• smoke_layout_detector.py: --allow-no-table flag; doc/output cleanup
• smoke_structure.py: dedup now default (no --dedup-bands flag needed)

notebooks/04_phase2_layout.ipynb:

• Step 3a syntax fix; Step 5/6/7 text updated to reflect final results
• Step 7d annotated: validator fix -> expect 285 OK / 1 WARN

docs/:

• DEVLOG.md: Phase 2 MVP finding (2026-06-02); unblocked from .gitignore
• PLAN.md: Phase 2 status updated; fixed DocLayNet subset eval framing
• README.md: Phase 2 layout/crop metrics added
• DESIGN_SPEC.md: fallback rule + dedup_row_col_bands documented
• .gitignore: remove DEVLOG.md exclusion

[AD2000X]

Summary

• Layout detection: Aryn/deformable-detr-DocLayNet primary, table_threshold=0.30, dedup_iou=0.70; fallback fires only when primary finds ≥1 table but none above threshold
• MVP gate (seed=42): recall@0.50=0.900, precision@0.50=0.916, FP crop rate 6.5%
• Band dedup postprocess: dedup_row_col_bands (1-D NMS) now default in normalize_tatr_prediction; fixes TATR overlapping row/col bands on dense crops
• Structure handoff smoke (seed=42, n=286): 285 OK / 1 WARN (0.35%) — both Phase 2 gates passed

Test plan

• pytest — 219 passed
• Step 6b/6c: IoU + FP gate passed on Colab T4
• Step 7d: structure smoke 285/286 OK (0.35% WARN, well under 5% gate)