[erlis]

Site 2 of the split agreed on #681, coordinated with @brett19 on #717. It takes off the loop-bisector known mismatches that #717 and #718 parked.

What. SmFace::ClassifyLoops' loop bisector (sms_FindAndClassifyLoopBisector) ran IsSeam on a private b.clone() and each CurveOnClassify on a fresh snapshot. Upstream runs both on the face's own B-rep: IntersectTSBound → ComputeUVPoint (bOKToMakeUVTrimCurve=TRUE) makes the missing UV trims, and SetUVTrimCurve resets the vertex/face gaps. In the port those edits were lost. Now:

• The bisector's IsSeam and classifications share one CurveOnClassifyCache snapshot, threaded through classify_loops and SetProps_Stage3 as crate-private _cached variants (CurveOnClassifyCache::snapshot gives IsSeam the same snapshot). The public classify_loops, set_props_stage3 and find_and_classify_loop_bisector(_into) keep a private cache, so they behave as before.
• The healer's in-place SetProps (set_props_with_query_gaps) runs the seam classifications and ClassifyLoops on that one snapshot. It copies the snapshot's UV trims, tolerance growth and gap caches back with apply_uv_trim_updates, as IsInnerShell does since fd54ae4. set_props_with_seam takes the cache instead of a seam closure, so both stages share it.
• HealData::cache_face_props_stage3 takes &mut Brep and uses that in-place SetProps; its callers already hold the B-rep mutably.

Known mismatches. Of the 49 entries listed for this cause:

• 37 removed: heal.heal_data_p1 28 and topology.brep_p1 9 now match C++ bit for bit, including port: topology.face_p4 (stacked on #693, merged with main) #718's file_plane_circ, file_plane_circ_rebuild_reverse and plane_circ_heal.
• 12 kept (topology.brep_p10), with an updated reason and a new plan/findings.md entry. They take the in-place path: instrumented, gen0165 and read_Fillet_bottle reach set_props_with_query_gaps Stage3 → classify_loops_cached. Their first differences are unchanged except gen0040's. Reading the bisector's UV guess on the snapshot changed none of them, so their cause is something else and I haven't traced it. This includes the three with port: topology.face_p4 (stacked on #693, merged with main) #718's HasSeamProblem caveat.

Difftest (full sweep, main 92454ca vs this branch): 0 failed on both sides. 37 cases go known mismatch → pass, with no other transitions.

28 heal.heal_data_p1 | known_mismatch -> pass | e.g. box_stage2_stage3_done_op_19
9  topology.brep_p1  | known_mismatch -> pass | e.g. Mouse_heal

base:   PARTIAL heal.heal_data_p1 cases=495 passed=456 failed=0 deferred=11 max_rel_err=1e0 known_mismatch=28
branch: PARTIAL heal.heal_data_p1 cases=495 passed=484 failed=0 deferred=11 max_rel_err=0e0
base:   PARTIAL topology.brep_p1 cases=2627 passed=2587 failed=0 deferred=15 max_rel_err=1e0 rejected=16 known_mismatch=9
branch: PARTIAL topology.brep_p1 cases=2627 passed=2596 failed=0 deferred=15 max_rel_err=0e0 rejected=16
base:   PARTIAL topology.brep_p10 cases=1404 passed=1380 failed=0 deferred=4 rejected=7 known_mismatch=13
branch: PARTIAL topology.brep_p10 cases=1404 passed=1380 failed=0 deferred=4 rejected=7 known_mismatch=13

(The remaining deferrals in those units are blocked by other units.)

Checks (VM): fmt; clippy with and without difftest/cpp; cargo test -p solidcore (2126 passed); wasm; cargo test --profile difftest -p difftest --features cpp; the full sweep.

🤖 Generated with Claude Code

[bherbruck]

Approved by the automated first-pass review: CI green, mergeable, no CI/build/dependency file changes, no red-flag patterns in added code (process/network/env/unsafe outside FFI/blobs/hidden unicode), 0 known-mismatch entries added. Logic is gated by CI and the difftest merge queue.