Rebuild ExcelTableCNN: working TableSense-inspired detector with PBR, grid-context backbone, and measured results (v0.3.0)

[Flagro]

Why

Neither of the repo's two model implementations ever worked: one computed no
RPN loss (loss collapsed to ~0 while learning nothing), the other crashed on
construction, and the data layer produced zero-area ground-truth boxes and
constant border features. No tests, no packaging, no CI.

What

• One model path: torchvision Faster R-CNN based; legacy + experimental
  code archived; package restructured into data/model/training/evaluation.
• Fixed data pipeline: 30-channel featurization (paper's scheme) with
  parity-tested openpyxl and native xlrd backends — no LibreOffice
  needed for .xls; half-open box convention; file-hash tensor caching;
  MD5-verified corpus downloads.
• Detector upgrades: corpus-tuned anchors (GT coverage @IOU≥0.7:
  45%→71%), PBR boundary snapping (per-edge offset classification →
  cell-exact boxes), and a novel grid-context backbone (row/column
  density priors + axial strip pooling — not in TableSense).
• Tooling: excel-table-cnn-train/eval/detect CLIs, per-component loss
  logging, device=auto (CUDA→CPU; MPS tested, opt-in), EoB-0/EoB-2
  metrics, uv packaging (uv.lock, no requirements.txt), CI.
• 106 tests, incl. the keystone gate: overfit one sheet to EoB-0.

Results (240 train sheets, 40 epochs, CPU, held-out 30 sheets / 67 tables)

Configuration	EoB-0 R/P	EoB-2 R/P
v1 baseline	0 / 0	23.9% / 41.0%
+ paper levers (30ch, anchors, 14×14)	0 / 0	19.4% / 24.1%
+ PBR	10.4% / 14.3%	29.9% / 40.8%
+ grid-context (full)	16.4% / 25.6%	40.3% / 62.8%

Paper reference: 91.3/86.5 EoB-2 — on 25× more training data. Next lever:
full-corpus GPU run (notebooks/train_kaggle_colab.ipynb).