test: pin different-net trace crossings per bug report

[DPS0340]

Why

Nearly every bug report in tests/bug-reports/ is verified by an SVG snapshot. That catches change, but not quality — a trace crossing that appears or disappears is a handful of pixels in a large regenerated SVG, and easy to approve without noticing.

Different-net crossings are the thing several of these reports are actually about, so I measured them across every imported board:

crossings	segments	board
11	149	bug-report-20260706T213649Z
6	199	bug-report-20260706T220324Z
5	168	bug-report-20260721T221026Z
4	52	bug-report-20260716T144856Z
3	38	bug-report-20260717T022934Z
3	67	bug-report-20260707T134549Z
2	20	bug-report-20260708T055430Z
2	154	bug-report-20260707T092615Z
2	28	bug-report-20260707T230831Z
1	80	bug-report-20260707T140410Z
1	22	bug-report-20260707T020342Z
0	—	(8 further boards)

40 crossings across 19 boards. None of that was visible from CI before.

What this adds

• tests/fixtures/traceCrossings.ts — getTraceCrossings(traces), mirroring the shape of the existing getTraceLabelCollisions helper next to it. Counts only proper crossings between segments on different nets; shared endpoints on the same net are excluded, since traces legitimately meet there.
• tests/different-net-trace-crossings.test.ts — pins the count per board, one named test each.

These numbers are the current state, not a target. Several of these boards are open bugs and the counts should come down. The value is that they can't quietly go up: a routing or cleanup change that makes a board worse now fails an assertion that names the board, instead of sliding into a regenerated snapshot.

There's also a guard test asserting the pinned set matches what's on disk, so a newly imported bug report can't go unmeasured.

Verification

• The guard actually bites. Changed the expected count for bug-report-20260706T213649Z from 11 to 10 and got Expected: 10, Received: 11. A pinning test that can't fail is worse than no test.
• Full suite: 164 pass / 0 fail / 4 skip (was 144 before, +20 from this file).
• tsc --noEmit clean, biome clean.

One implementation note: I initially read traces from netLabelNetLabelCollisionSolver, the last pipeline step — but its getOutput() returns only netLabelPlacements. tsc caught it. Final trace geometry comes from traceCleanupSolver2, which is what this reads now.

Follow-up

If this is useful, the natural next step is fixing the worst boards and lowering the numbers in the same PR. bug-report-20260706T213649Z at 11 is the obvious first target — happy to take that on.

Diff: 2 new files, no existing file touched.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
schematic-trace-solver	Ready	Preview, Comment	Jul 26, 2026 6:13am

[DPS0340]

Updated this to pin overlaps as well as crossings, because I found the one-sided version scoring a good change as a bad one.

I had used the earlier version of this counter to report a "regression" from #727. It was not a regression. Counting different-net collinear overlaps too:

                        crossings   overlaps
ca7e80c (before #727)       40          1
be20aa7 (#727 merged)       41          0

Same pair of segments on both sides, on the single board that moved:

bug-report-20260707T092615Z
before:  overlap   C_SENSE.2-C1.2  ||  available-net-orientation-13-BAT_POS  (vertical, span 0.141)
after:   crossing  C_SENSE.2-C1.2  ×   available-net-orientation-13-BAT_POS  (x=-9.809, y=-1.764)

#727 separated two nets that had been drawn on top of each other and left them crossing at a point instead — the right trade, since an overlap looks like one trace and hides that two nets are there.

That is the argument for pinning both. A crossings-only guard is gameable in one specific direction: any crossing can be eliminated by making the two nets run along each other, and a one-sided counter reads that as an improvement. Pinning both makes the trade explicit — a PR that moves a board from 2/1 to 3/0 has to say so, rather than passing or failing depending on which quantity happened to be measured.

Still measurement-only — no lib/ changes:

• getTraceOverlaps added next to getTraceCrossings in tests/fixtures/traceCrossings.ts. Shared endpoints are excluded, since traces meeting at a point is normal; only a shared span counts.
• the per-board table now pins { crossings, overlaps }.

Verification, since a guard that cannot fail is worth nothing:

• checked out ca7e80c with the new fixture → 092615Z fails on overlaps: 1, the other 19 pass. The counter detects the real state the crossings-only version missed.
• current main → all 20 boards pass.
• full suite 167 pass, 4 skip, 0 fail; bun run format:check clean; bunx tsc --noEmit exits 0.