audit YOLO geometry and overlapping validation rows

docs/CODEX_EXECUTION_LOG.md

@@ -12595,3 +12595,46 @@ Open:
 - This is explicit AI-assisted triage, not human adjudication. It improves the
   evidence and review precision but does not unlock governed production
   training or checkpoint promotion.
+
+## 2026-08-09 - Row-level geometry and cross-tile independence audit
+
+### Improved
+
+- Added a reproducible, read-only label-row auditor for sub-threshold pixel
+  dimensions, extreme aspect ratios and tile-edge clipping, with separate
+  category manifests accepted directly by the visual QA renderer.
+- Extended the renderer with orange extreme-aspect, blue small-dimension and
+  green tile-edge highlighting while preserving the existing duplicate/nesting
+  colors and source row indices.
+- Added an exact cross-tile repetition auditor that reconstructs global pixel
+  boxes from the exporter's `_r<row>_c<column>` offsets. Edge-clipped rows stay
+  explicitly unlinked rather than being guessed back to source objects.
+- Made checkpoint-matrix output disclose dataset tile overlap and explicitly
+  state whether validation rows are independent whenever a dataset summary is
+  available.
+
+### Tower findings
+
+- Reviewed all 79,192 rows: 1,812 have a dimension below 4 px, 26 have aspect
+  ratio at least 8 and 5,079 touch a tile edge; 6,657 unique rows are involved.
+- Inspected the complete 23-tile extreme sheet, all four small-dimension pages
+  covering 198 tiles, and a 64-tile edge sample. Most extreme shapes are valid
+  elongated structures, but 16 of 26 also fall below the 4 px floor.
+- Reconstructed 34,082 unique interior objects from 74,113 interior label rows.
+  Tile overlap adds 40,031 exact repeated rows across 24,701 object groups,
+  with maximum repetition four and zero cross-split groups.
+- Froze the next-corpus decision: restore `min_label_px >= 4`; test any stricter
+  edge-visible ratio as a separate immutable ablation rather than modifying
+  this corpus in place.
+
+### Verified
+
+- Focused auditor, renderer and checkpoint-evidence suite: 18 tests passed.
+- Scoped Ruff check, format check and Git whitespace check passed.
+
+### Remaining limitations
+
+- Existing overlapping-tile mAP is valid for relative non-protected ranking,
+  not independent object-level accuracy. A future benchmark must aggregate or
+  deduplicate objects before uncertainty claims.
+- No source label, frozen dataset, model weight or live runtime was changed.

docs/TODO.md

@@ -1159,6 +1159,14 @@ This file now starts with the current implementation status. Older preparation/b
   contact sheet; AI follow-up found no systematic duplicate-label pattern.
 - [x] Bind the relationship audit back to exact label-row indices and render
   all 86 implicated rows with type-specific highlighting for focused review.
+- [x] Audit all 79,192 labels at row level for sub-4-pixel dimensions, extreme
+  aspect ratio and tile-edge clipping; inspect the complete small/extreme
+  sheets and retain immutable hashes.
+- [x] Reconstruct exact cross-tile interior objects and prove that overlap
+  repeats 40,031 label rows but causes zero reconstructed train/val crossings.
+- [ ] Build the next immutable experimental corpus with `min_label_px >= 4`
+  and compare the existing 0.35 edge-visibility policy against a separately
+  versioned stricter ablation before any new governed training.
 - [ ] Convert the AI-assisted ledger into no stronger claim than experimental
   triage; a real human must independently review and sign the frozen artifacts
   before the governed training wrapper may unlock.

docs/reviews/2026-08-09-ai-assisted-checkpoint-review.md

@@ -23,6 +23,15 @@ The checkpoint matrix used only the 36-image non-protected validation split:
 measured mAP50 `0.345426` and mAP50-95 `0.140719`. The `reviewedexp6`
 challenger measured `0.368177` and `0.155438` respectively.
 
+A later exact cross-tile reconstruction established that this 512 px corpus
+uses stride 256. Across the full corpus, 74,113 interior label rows represent
+34,082 unique reconstructed objects; 40,031 rows are overlap repetitions and
+one object can occur four times. No reconstructed object crosses the
+train/validation boundary, but the tile rows are not statistically independent.
+The matrix therefore remains useful only for relative non-protected candidate
+ranking. Its mAP values are not an independent object-level accuracy estimate
+or a basis for narrow confidence claims.
+
 At confidence `0.25` and match IoU `0.25`, the active model measured F1
 `0.633058`; the challenger measured `0.642599`. Both produced zero detections
 on the 18 pure-background validation tiles at this threshold. At confidence
@@ -129,3 +138,26 @@ rewrite candidates. Excluding the 36 complete tiles would also discard 17,167
 unflagged labels, so no automatic tile exclusion or label mutation was made.
 The checksum-bound highlighted sheet and summary are recorded in the AI review
 ledger.
+
+## Row-level geometry and overlap follow-up
+
+The complete 79,192-label corpus was additionally audited at label-row level.
+The immutable manifest identifies 6,657 unique rows with at least one geometric
+training-risk signal: 1,812 rows have a width or height below 4 pixels, 26 have
+aspect ratio at least 8 and 5,079 touch a tile edge. The categories were
+rendered separately. All 23 extreme-aspect tiles and all 198 small-dimension
+tiles were inspected, plus a 64-tile edge sample.
+
+The extreme-aspect group predominantly shows plausible elongated sheds and
+building components. Sixteen of its 26 rows are also below 4 pixels, allowing
+the resolution floor to address most ambiguous extremes without deleting valid
+long structures. The small-dimension sheets contain many visually marginal
+miniature targets. The next immutable experimental corpus should therefore use
+the exporter's normal minimum dimension of at least 4 pixels instead of this
+legacy corpus's 3-pixel override. This finding does not justify mutating the
+frozen corpus or retroactively changing its checkpoint.
+
+The edge sample shows expected clipped buildings under the 0.35 minimum-visible
+policy. Raising that value may reduce partial-target pressure, but it must be a
+separately versioned ablation because removing all 5,079 rows without checking
+the original visible fraction would be unsound.