Merge pull request #2083 from tangency-labs/plan/smlib-fork-port-audit

plan: SMLib fork port audit record (2026-10-31)

plan/findings/general--smlib-fork-port-audit-2026-10-31.md

@@ -0,0 +1,86 @@
+# General: SMLib fork commits not cited by solidcore, and what was measured on them (2026-10-31)
+
+**Unit:** general (the opt-in `Kernel::fixes` flags that port fixes from the SMLib fork).
+
+## Scope
+
+- **Fork:** tangency-labs/smlib, all 28 branches as fetched on 2026-10-31 at 18:30 UTC. The fork's `main` head was 83748ba9876902dd8e91ef72849dedc1a2f1dcb9 (2026-10-28).
+- **Commits:** every commit past the 1.0.1 base 3fd5b95 that touches `source/SMLib` or `source/SM_API`. That is 174 on the fork's `main` and 137 more on its other branches, after dropping branch commits whose subject matches one on `main`: 311 commits. Only SMLib and SM_API were read as port sources; `capi/` was read for test names and test inputs only.
+- **Solidcore:** checked against main 46b22403cd (2026-10-31) and the 53 PRs open then (titles, bodies and diffs). A fork sha counts as cited when a 7- to 40-hex token anywhere in crates, plan, docs, fixtures, scripts or harness is a prefix of it. The measurements below ran on main 68a6b5a00d.
+- **SolidCraft's preset:** `fixes()` in `crates/smlib/src/core.rs` at tangency-labs/solidcraft#8526's head, brett19/solidcraft 0b627d637325c67f07d943d812bf0e58d4647b70 (blob 68d628c3). It sets 188 flags on. Its `HELD_BACK` list (33 flags) and the flags newer than its solidcore pin stay off. Each probe ran this preset, the default kernel, and every flag on (labelled "not the preset").
+
+## Counts
+
+185 of the 311 are cited by sha. The other 126 sort as follows:
+
+| Section | Count | Meaning |
+| --- | --- | --- |
+| A | 31 | Unported behavioural fixes: no flag or record port covers them (statuses below). |
+| B | 10 | Status hygiene: not ported, by the ruling of 2026-10-29 cited in the WR-260 and WR-562 addenda. The sites that change results are behind `query_statuses_pass_on`, `fillet_statuses_pass_on`, `brep_edit_statuses_pass_on` and `sweep_statuses_pass_on`. |
+| C | 26 | C++-only, nothing to port: build fixes, memory and ownership bugs that arena ids rule out, crash guards the port already has, diagnostics, test instrumentation. |
+| D | 53 | Covered by an existing flag or record port that does not cite the sha (bit-for-bit results below). |
+| E | 6 | Cited only in PRs open at the time: 158cbe7 and 1612a7d (#2061), 1ff9264 (#1913), 5fbffc9 and 6eecf04 (#1828), 8843967 (#2066). #2061 and #2066 have since merged. |
+
+**F: cited, but only partly ported.**
+- d6a6d53: subsumed by #1580's branch, so not ported as written.
+- 47507b6 and 8088f5e: the rest of each is status hygiene.
+- a759622: its `SmBrep::LocalOperation` status change is not ported.
+- a96ce9f: covered by WR-615's `offset_faces` API, but whether that covers all of a96ce9f is not checked.
+- 0b4de8c: ported as WR-56's flag; not yet on the fork's `main`.
+
+## Section A
+
+**Measured: no port.** Each passes under the preset on main 68a6b5a00d, and the default kernel still shows the C++ defect.
+
+- **c340962** (#429, box minus a cylinder at dx 0, 3e-4, 3e-5 and 3e-3).
+  - **Preset:** every dx gives a closed manifold solid with no Level-2 failures.
+  - **Default:** dx 3e-4 is not manifold, with 6 Level-2 failures; dx 3e-5 fails in `make_faceuse` with 1001.
+  - **Volumes** against the closed form (2.25 minus the disk-square overlap) × 7.1:
+    - dx 0, 3e-4 and 3e-3: within 3e-9 relative;
+    - dx 3e-5: 3.4279547 against 3.4279462, 2.5e-6 relative. The circle passes 3e-10 from the box corner there.
+- **f02ab6e** (#473, the holed cube of `tests/fillet_wr300_320.rs`).
+  - **Preset:** chamfers 6 and 9 give closed manifold solids of 3650 and 3350, with no Level-2 failures.
+  - **Default:** volume 4000, not manifold.
+  - **Chamfer 4:** 3840 in every leg.
+- **1b8c9bc** (#422, the block of `tests/issue_422_round_over_through_hole.rs`).
+  - **Preset:** R4 gives a closed manifold solid of 7580.29, with no Level-2 failures.
+  - **Default:** not manifold.
+- **7ee0876** (the fork's own sphere-cap thicken, as built by `tests/issue_1502_z5_cap_thicken.rs`).
+  - **Preset:** both rim shapes, one closed circle and two open half circles, give manifold, Level-2-clean solids of 346.6223894574, against 331π/3 = 346.6223894461.
+  - **Port:** `offset_closed_curve_trim_arc` (WR-463, the fork's 34eaa6b) covers it. It lacks this commit's own second-trim-point test, but that test is not needed for this case.
+
+**Reconstruction only.**
+- **9fd7e7e** (rig booleans/562). The exact operands exist only in the fork. On a reconstruction from the commit's description (a 60 × 40 × 10 box with its vertical edges rounded r3, united with a 60 × 40 × 1 prism of the same rounds), every leg, the default included, gives a closed manifold solid with the closed-form volume. A pass proves nothing about rig booleans/562.
+
+**Unreachable.**
+- **e309858** (#390). `SmApiTessellate` (`SmApiBrep.cpp:679`) calls `SmBrep::ConvertToPolyBrep`, which runs `SmTess::DoTessellation` (`SmTess.cpp:12316`), where the fix runs after triangulation. None of the three has a Rust counterpart or a unit in `harness/callgraph/units.json`. `query::tess` ports only `SmTessSrfCache` and the tessellation drivers.
+
+**On side branches, dated 2026-10-29 to 2026-10-31, still moving.**
+- blend-end-event: 55bffe0, 922f333, b6f88e4, e881cdb, plus the G2 tangent-end series 6ca3405, 68f6913, 6d73e4f, b34d61b and a87c502, with its revert 0cc48bd.
+- smlib-curved-setback: 124e33d, a47424b, b3a33fe.
+- blend-exact-classes: 1866b03, 344b305, 808c7f6, c290e61, f473917.
+- smlib-blend-refusals: f63fe26.
+
+**Not yet probed.** On the fork's `main`: 3545008, 8c4bda5, 8ea6a20, 9a160ba, f653ac9 and 1279539. f653ac9 is a new API (per-edge second offsets).
+
+## Section D: bit-for-bit
+
+For each of the 53 entries, the commit's SMLib diff was compared with the Rust code at every site that reads its flag. Nothing was run.
+
+**Outside blend and offset** (34 commits: topology, boolean and geometry):
+- 19 exact;
+- 11 partial;
+- 4 fixed differently. These four are 7a6b4ad, f72c79b, 69d8857 and 7a8f4d3, recorded in #2079.
+
+**Blend and offset** (19 commits):
+- **15 exact.** Six of these differ only on error paths or leave one point unchecked.
+- **2 partial:**
+  - 7ee0876: moved to section A (above).
+  - 9504880: its "on target" test uses geometry plus `MARK_MOVED`, where the fork uses a surface mark. This deviation is documented in `crates/solidcore/src/offset/offset_faces.rs`.
+- **1 superseded:** 2bc03c6, by a15b5a2, which WR-462 cites.
+- **1 covered:** 1ccd2af, whose changes are all in e418750, which WR-665 cites.
+- 2412a80's fix is in our default path. Its flag `temporary_radius_setup_errors` also fails the constructor, which the fork's does not. That is solidcore's own extension, and SolidCraft holds the flag back.
+
+The probes of the partial entries are pending. No outcome is claimed for them here.
+
+tangency-labs/solidcraft#8526 comment 6548297899 asks Brett whether SolidCraft wants the fork's version of the four fixed-differently commits (7a6b4ad, f72c79b, 69d8857, 7a8f4d3).