Cape31 boom-cam 2026-09-02 — read-only baseline sweep (SAIL-145)

568 curated frames, boom-mounted Pi cam, tack-normalized (sail at bottom). Lines drawn: v26 deep20 sam3. Sorted worst-first: 0.5*max(0, 3-min(v26Count, deep20Count)) + (chamferV26Deep20 or 0); chamfer=symmetric mean NN distance, normalized coords, 50-pt resampled polylines; null chamfer when either model has 0 stripes. Soft = sail-band variance-of-Laplacian ≤ 15 (stays in the sweep; barred from future test slices only).

v26 vs deep20 chamfer (normalized units): mean 0.01998 · median 0.02368 · p90 0.02627 · max 0.06721 over 564 frames. Caveats: sam3 line counts include non-draft-stripe lines (it also traces the sail foot), so its row is not stripe-count-comparable — analysis.json's sam3CoverageOfV26 is the matched metric. Most cards after the first screenful are deep20-two-stripe frames (via THIS preview config it consistently drops the bottom stripe on boom-cam framing; the same weights at the SAM3 pod's finder settings do not) — the wall of 2-counts is the finding, not a rendering bug.