feat(0125-TorsoChestBackLandmarkSpecsCaptureCompare): torso landmark capture, compare, and placement consume

[Ryan-AI-Studios]

What

Product path for torso/chest/back form-read specs: frozen assist ids → fuse XYZ (left Y overlay; back X/Z) → sidecar torso_metrics.json → CLI proportion blockout-torso-compare (photo vs RECIPE vs optional scene) → agent-facing deltas. MCP catalog 48→49 (mesh_proportion_blockout_torso_compare). Soft new-id placement only: measured scap/mid_back Y/Z when finite; otherwise 0105/0090/0066/0074/0118 priors.

Why

Package A has bust/waist + chest/breast depth, not sternum/scap/mid_back. Agents were guessing from H-frac priors. Dual-source compare is authoring QA only.

Honesty

TORSO_COMPARE_HONESTY = proportion_torso_compare_not_mesh_or_print_success. Green compare is not mesh or print success. No remake/D7 (Package A lacks new ids → emit identical). Report stay 1.2.0. Recipe stay 1.4.0. No Pose/Holistic pin. No DEPTH_PAIRS consume into oval/breast consts.

Tests

tests/test_proportion_torso_landmarks.py A1–A6 + back X/Z + left Y. tests/test_proportion_torso_compare.py B/C/D/F + sidecar unlink + hold_priors on frozen knobs. tests/test_proportion_torso_soft_consume.py E1–E5 + B32. Catalog pin-retarget 48→49.

---

Note

Medium Risk
Touches multi-view XYZ fuse and RECIPE scap/mid-back placement, so bad landmarks can shift blockout geometry. Scope is authoring QA with fail-closed dumps and no auth/data-path changes.

Overview
Adds track 0125 torso/chest/back form-read: new assist ids (sternum_mid, costals, scap/mid-back), fuse XYZ from left (Y) and back (X/Z) without inventing missing axes, and authoring QA blockout-torso-compare (CLI + MCP; catalog 48→49).

Compare writes torso_compare.json plus optional sidecar torso_metrics.json (report stays 1.2.0). Roles get signed mm deltas, form-read tokens (three-tire, saucer plates, breast gap, clav pride), and a closed suggestion set. Frozen oval/breast knobs stay hold_priors; only finite scap/mid-back Y/Z overlay RECIPE plate centers after existing equalize (no abs of measured Y).

Honesty: TORSO_COMPARE_HONESTY — not mesh or print success. No new DEPTH_PAIRS and no retune of 0105/0090/0066/0074/0118 consts.

^{Reviewed by Cursor Bugbot for commit c5df10a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit c5df10a. Configure here.}

[cursor]

Stale chest-derived torso metrics

Medium Severity

build_torso_metrics computes breast_rear_vs_chest_front_m and chest_clav_pride_m from recipe chest-front Y, then later overwrites chest_front_y_m with measured chest_front without refreshing those derived values. On packages with left depth, breast_disconnected / chest_clav_pride form-read and the sidecar metrics disagree with the reported chest-front Y.

 

^{Reviewed by Cursor Bugbot for commit c5df10a. Configure here.}

[cursor]

Measured past uses center not surface

Medium Severity

When measured scap/mid_back Y is present, scap_rear_past_m / mid_back_rear_past_m subtract chest/waist rear surface from the landmark Y. Soft consume writes that same Y onto the plate center, while the recipe-only path measures past from center plus ry_m. Measured past is therefore short by about plate ry, so compare metrics disagree with placement semantics.

Additional Locations (2)

• src/meshops/proportion/blockout_recipe.py#L1967-L1991
• src/meshops/proportion/blockout_recipe.py#L2178-L2202

 

^{Reviewed by Cursor Bugbot for commit c5df10a. Configure here.}