fix(router): window repair searches and commits at the same width, and merges via barrels

[ecto]

What changed

Scope changed after #681 landed. This PR originally moved mutual-legality proving and per-net class widths into route_window_complete itself. #681 then implemented both caller-side — resolve_cluster probes-then-commits per path in order, and validate_and_commit gets the same property for free via &mut RouteSession. Caller-side is the better place: the caller probes the real session, whereas a router-side replay can only probe a scratch clone that cannot see copper committed outside the window. So that work is dropped as superseded, and this PR is now just the part #681 left behind.

#681 fixed the CM5 verdict driver. It did not fix auto.rs::joint_window_repair, the other consumer of the window router, which had three gaps that each cost connections by finding a routing the commit then rejects:

gap	consequence	fix
searched at board-default width, committed at session.width_for(net, width)	a corridor found at 0.08 mm is rejected at commit when the net's class is 0.2 mm — the connection is lost to a search/commit mismatch, not to the board	search at the widest class width among the window's nets
derived vias from a naive windows(2) layer-change scan	F.Cu → In1 → In2 at one point emitted two coincident vias, stacking drills at zero spacing and failing hole-to-hole against itself	new path_vias, which merges a run of transitions at one point into a single barrel
passed no via geometry	pad probed at a heuristic radius, drills never checked	pass a real ViaClass { pad_diameter, drill } so the router's legality model is the commit rule

path_vias is pub in complete.rs because the merging rule belongs next to the emitter, not in a caller — #681 wrote a copy of it inline in the example, and that copy can now delegate.

Tests

• a_run_of_layer_changes_at_one_point_is_one_barrel — F.Cu → In1 → In2 without moving collapses to one F.Cu → In2 barrel. This is the bug that produced self-colliding drills.
• separate_layer_changes_stay_separate_barrels — guards the other direction: two transitions at different points stay two vias, and a layer change whose endpoints don't meet is not a via at all.

cargo test -p vcad-ecad-pcb --features gpu → 266 passed, 0 failed. cargo fmt --all --check clean. cargo clippy -p vcad-ecad-pcb --features gpu --lib --example cm5_verdict -- -D warnings clean.

Reviewer notes

A fourth gap is left open, deliberately. auto.rs still passes an empty terminals slice, so its stubs can surface off the pad's own layer — the net-island failure TerminalLayers exists to prevent. Closing it needs the example's terminal_layers (~45 lines, depends on point_on_segment and Driver::span) lifted into the library so both callers share one implementation. That is a larger change than this fix and wants its own PR; I did not want to half-implement it, since a too-narrow layer set fails closed but silently loses routes.

No CM5 before/after numbers this time. The earlier revision of this PR reported them, but that measurement was against the pre-#681 router and no longer describes this code. The change is argued from the three mismatches above and covered by unit tests rather than a board-level count.

Cargo.lock is intentionally untouched. Building against current phyz (main, 0.3.1) wants a 59-line lock update, because vcad main still pins phyz 0.3.0 while CI clones phyz's default branch. That refresh is unrelated to this fix and CI re-resolves the lock itself (no --locked on the workspace build), so it is left for a dedicated change.

🤖 Generated with Claude Code

[chojiai]

Choji has used up today's reviews

Your team has reached the free plan's limit of 5 reviews per day, so Choji stepped aside on this pull request. That's a plan limit, not a reflection of the change itself.

Your free reviews reset tomorrow. To have Choji review every push without waiting, upgrade to Pro for unlimited reviews.

_{You're on the free plan · 5 reviews/day.}

[vercel]

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

4 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
mecheval	Ignored		Jul 25, 2026 1:39pm
vcad	Ignored		Jul 25, 2026 1:39pm
vcad-docs	Ignored		Jul 25, 2026 1:39pm
vcad-mcp	Ignored		Jul 25, 2026 1:39pm