[lightningpixel]

Closes #298

Summary

Adds repair_mesh, a composite op driven by the inspect_mesh diagnostic (T-A2): "repair this mesh" now turns a holed, non-manifold mesh into a watertight, manifold one — checked by re-running inspect_mesh on the output, not just by the repair's own bookkeeping.

The existing repair op could not meet that bar: even with the params suggest_fixes chose, it left holes open (size limit), introduced inconsistent winding, and had no way to fix flipped or inside-out faces.

This PR also fixes a dependency bug that already affects inspect_mesh on dev — see below.

How repair_mesh works

1. Diagnose the mesh (position-merged, so UV seams are not mistaken for holes).
2. Run only the steps the diagnostic calls for: remove degenerate faces, fix non-manifold edges, fill holes (no size limit), re-orient faces.
3. Re-diagnose; repeat up to 3 passes, stopping early once the mesh is closed and manifold or a pass makes no progress.
4. Flip the mesh if it ends up inside-out.

Each pass is judged on position-merged geometry, exactly as inspect_mesh sees it — an earlier version judged pymeshlab's raw output and reported success on meshes that still had non-manifold edges once re-inspected.

• A sound mesh is left alone: no file is written.
• fill_holes: false (for meshes meant to stay open) only runs face-preserving fixes, since removing faces would open holes it can't close.
• The result reports before/after topology, the remaining issues, and textures_dropped — the output is geometry only.

Other changes

• inspect_mesh now flags inside-out meshes (is_inside_out, counted against printable).
• suggest_fixes proposes repair_mesh, so inconsistent winding becomes fixable.
• The generic POST /optimize/op/{id} route writes its output next to the input under a unique name ({stem}_{op}_{id8}.glb), like the dedicated edit routes, instead of Workflows/.
• Agent: repair_mesh tool (reports what remains and warns when textures were dropped). MCP: modly_repair_mesh.
• The legacy repair op is unchanged — the mesh-repair workflow node depends on it.

Dependency fix (affects dev today)

The shipped venv installs only api/requirements.txt (50 packages, verified with a pip dry-run), which does not include scipy — trimesh treats it as optional. inspect_mesh and suggest_fixes (T-A2) need it for connected parts, hole loops and vertex merging, so they fail in the installed app; tests passed only because dev venvs get scipy through torch.

• scipy>=1.10 is now declared in api/requirements.txt (the requirements hash change re-runs pip once on update).
• repair_mesh no longer uses trimesh.repair.fix_winding (needs networkx, also undeclared); it re-orients with pymeshlab.
• New guard test: runs every mesh-op family with networkx unimportable, including the re-orientation path, and checks scipy stays declared. Verified to fail on the networkx-based version.

The pre-push hook caught the networkx dependency: it failed on the system Python, which has no networkx.

Testing

• npm test — 275 Python tests (4 skipped), 262 Node tests, all passing; also run on a Python without networkx.
• 11 new repair_mesh tests: holed + non-manifold → watertight & manifold, large hole, flipped faces, inside-out, sound mesh untouched, fill_holes: false never removes faces, textured input, input untouched, no-triangle file rejected.
• Khronos samples: Avocado → watertight; DamagedHelmet 83 holes / 7 non-manifold edges → 1 / 1, reported as remaining.
• ~9 s per pass on a 1.2M-face holed mesh.

Known limits

• Textures and UVs are dropped (as with the legacy repair); keeping them is a follow-up.
• Self-intersections and disconnected parts are not handled.
• Not yet run on real TripoSR / Hunyuan outputs.