Skip to content

FINDINGS-INT.md

Latest commit

 

History

History
133 lines (101 loc) · 5.1 KB

FINDINGS-INT.md

File metadata and controls

133 lines (101 loc) · 5.1 KB
Raw

Integration findings — hl/integrate (2026-08-26)

Adjudication per LANE.md. Branch hl/integrate merges lanes D, A, C in full and lane B segmentation selectively.

Commits merged (prior to B)

Lane Commit Scope
D f1c6594 tests/tools/step_census.py, handle-lock ground-truth.json
A 9692bd4 regiondump truth, --diag, CMake half-link fix
C 8d3f78a planeCylSideContact + mesh-aware g1Tangent; cap/generator fallbacks

Lane B — merged (segmentation only)

Taken in full (scoped to coarseFusionBand: 500 ≤ nTri ≤ 1200, handle-lock band; corpus S02=412 and S13/S14≤20 stay on default gates):

  • adaptCoarseSegmentParams — thetaPlaneDeg≥15°, thetaCylHiDeg≥70°
  • evaluateCommit G2 — chord-sagitta + epsCylRing for ≤8-tris patches
  • evaluateCommit G3 — skip when arc span < 0.35 rad
  • evaluateCommit G5 — partial span 30° + g5Micro (R≥5 mm, span≥12°, nBands≥2)
  • estimateNBandsFromPatch — dihedral-median D2 fallback
  • commitPlanesA3 — coplanar UF merge (10° on coarse shards)

Files: src/refit_grow.cpp, src/refit_segment.cpp.

Lane B — stripped (adjudication)

Rejected item Rationale
stl2step.cpp open-shell accept when STEP contains cylinders Violates watertight contract; open shell never ships
refit_build.cpp J6 “coarse keep” (return built cylinders on J6 fail) Same — relaxation of J6 closure
planeNearPerpendicular, planeCapCylinder, capCircleOnCylinder Overlap C's principled side-graze + cap fallbacks; C wins
Looser pickIntAna accept (sewTol*100) Overlap C's mesh-aware tolerance path
Global kG5SpanPartialDeg 40→30 in refit_internal.hpp Replaced by coarse-only 30° in evaluateCommit
g1Tangent revert to epsPlane-only Would undo lane C
B's adaptCoarseSegmentParams for all nTri≤1200 Narrowed to 500–1200 band to protect corpus fixtures

Proof (a) — build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DSTL2STEP_BUILD_TESTS=ON
cmake --build build -j 8
→ exit 0, all targets built

Proof (b) — handle-lock TrueForm

./build/stl2step tests/diag/handle-lock/handle-lock.stl \
  -o /tmp/hl-integrate.step --engine trueform --quiet

RESULT (exit 2 — written with warnings, R2 revert):

smoothBuiltComponents=0  smoothRevertedComponents=1
smoothPlanes=0  smoothCylinders=0  smoothBuiltCylinders=0
openShells=0  watertight=true  facesAfterSmooth=434
warnings: … IntAna plane|cyl … ×5; J6: shell not closed freeEdges=49 faces=479 recover=2;
          analytic rebuild reverted on one component -- kept faceted

Segmentation does run (regiondump --diag):

segment comp0 ok regions=27 planes=12 cylinders=15 fillets=0 rejected=0
distinctRadii=10  (0.50, 3.79×2, 5×2, 5.75, 6.0, 9.02, 10×3, 11.70, 13.87, 16.69 mm)

J6 still cannot close (49 free edges at plane|partial-cyl chains). Honest R2 revert fires — output STEP remains faceted. No closure weakening applied.

Census vs ground truth:

python3 tests/tools/step_census.py --expect \
  tests/diag/handle-lock/ground-truth.json /tmp/hl-integrate.step
EXPECT FAIL: faces 434 vs 28; cylinder 0 vs 15; circle 0 vs 30
(post-revert faceted output — expected until J6 closes)

Target ground truth: 28 faces (13 planes + 15 cylinders), 30 circles.

Proof (c) — regression

Bisect (rm -rf build && cmake && ctest -R p2real_live at each commit):

Commit p2real_live
92c2905 PASS
f1c6594 PASS
9692bd4 PASS
8d3f78a PASS
a095f01 PASS (scoped B)

Offending change (reproduced): unscoped lane B refit_grow.cpp — global evaluateCommit G2/G3/G5 relaxations + unconditional A3 coplanar merge (hl/B @ 33514d6). With scoped segment only, unscoped grow still fails S13/S14:

S13 comp0: faces=12/16 exploded360=1 censusOk=False
S14 comp0: buildFaces=False

Root cause: B's coarse cylinder/plane gates must not run on corpus fillet fixtures (S13=16 tris, S14=20 tris). Fix: guard all B segmentation relaxations behind shared coarseFusionBand() (500 ≤ nTri ≤ 1200); reject B's nTri≤1200 adaptCoarseSegmentParams and global gate edits. Centralised in refit_internal.hpp so segment + grow cannot drift.

ctest --test-dir build
→ 27/27 passed

ctest -R p2real_live                              → 1/1 pass
ctest -R "p2buildtest|gates_smoke|corpus_engine"  → 13/13 pass
ctest -R "p2buildtest|gates_smoke"                → 12/12 pass

Remaining gaps (ranked)

  1. J6 shell closure on handle-lock — 49 free edges after 479-face analytic attempt; mixed analytic + partial-cylinder polylines need arc-length MakeEdge / follow-on P2 work. Segmentation now reaches 12+15 regions (was 0 built).
  2. Large bores missing — 20 mm / 30 mm radii not segmented (too few facets in arc; need wider growth through 90° caps).
  3. IntAna plane|plane / cyl|cyl / remaining plane|cyl misses — several chains still polyline after C fix (reg 18|17 tangent=false in regiondump).