recipe: no horizontal flip and full-resolution masks by default (winner of the 2026-09-28/29 comparison)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

README.md

@@ -297,7 +297,7 @@ python tools/ledger.py show            # the table; `render` writes runs/README.
 sbatch slurm/train.sbatch 2026-10-02_alc
 ```
 
-`tools/train_run.py` trains with the recipe from `src/training/train_yolo_seg.py`,
+`tools/train_run.py` trains with the recipe from `src/training/train_yolo_seg.py` (since 2026-09-29 with no horizontal flip and full-resolution masks, the settings that won the comparison on unseen gold frames; any setting can be overridden with `--set KEY=VALUE`),
 copies `best.pt` to `models/registry/<family>-seg-<tag>-<date>/weights.pt` (LFS) with
 `metadata.json`, measures the weights on `data/gold/` with the studio's evaluation
 code (`src/eval/`, vendored), and gates against the active model's

tools/train_run.py

@@ -31,8 +31,11 @@ sys.path.insert(0, str(ROOT))
 
 from tools import ledger  # noqa: E402
 
-AUGMENT = dict(scale=0.7, translate=0.15, mosaic=0.5, close_mosaic=10, fliplr=0.5, flipud=0.3,
-               shear=5.0, perspective=0.0)
+# Recipe since 2026-09-29: no horizontal flip (the camera never shows the pin on the left) and
+# full-resolution masks (mask_ratio 1), the combination that kept run A's detection with the
+# tightest masks; before that fliplr=0.5 and mask_ratio=4, as in the legacy train_yolo_seg.py.
+AUGMENT = dict(scale=0.7, translate=0.15, mosaic=0.5, close_mosaic=10, fliplr=0.0, flipud=0.3,
+               shear=5.0, perspective=0.0, mask_ratio=1)
 GATE_TOL = 0.005
 CLASS_TOL = 0.02
 GUARDED_CLASSES = ("loop_all", "pin")