fix(ecad): the SI receipt was measuring copper that is not a route

[ecto]

Re-measures the full-board CM5 vcad.si-claims/1 receipt after #684 and fixes
three fail-closed gaps it exposed. It does not reach all-claims-HOLDS, and
the reason is the point of the PR.

What the re-measurement found

The premise going in was that two claims needed closing. Measured on a fresh
full route, three of four were broken — vias_per_si_net had regressed through
its bound — and the headline skew turned out not to be skew at all:

/USB3-0.TX_P was being scored as a routed differential pair with 0.23mm of
copper across an 18.36mm pad span, while the DRC was already reporting it
UnconnectedNet. Comparing every leg's copper against its own straight
pad-to-pad distance — a ratio needing no connectivity model, so it cannot be an
artifact of how one defines an island:

board	legs	copper < 50% of span	min ratio
human	86	0	1.03
ours	84	7	0.01

The reported 38mm "skew" was essentially the length of whichever leg did
route. Compensation was never going to close it — there is no second leg to
match.

Three fixes

1. The pruner was defeated by shorts. dangling_copper_mask is documented
as dropping copper whose island touches no pad "of their net", but it judged
anchoring over the board-global touch graph and accepted any pad in the
component — so a dead island of net X that merely brushed net Y saw Y's pads and
survived. On a board with 845 shorts that is not a corner case. Now judged over
same-net connectivity, reusing the contact list the global pass already builds
(no extra broadphase); net ties still connect.

2. The receipt measured legs that are not routes. si_claims gated
"measured" on net_routed_length > 0, which sums any copper carrying the net
name. And coupled_fraction normalizes by P alone, so a stub beside a full twin
scores 1.000 — /MIPI1.D3_P read coupled 1.000 while carrying 38.52mm of
skew. The metric rewarded the failure. A pair is now excluded when either leg
holds less copper than its own pad span requires, and the exclusions are counted
and bounded by a new si_pairs_incomplete claim, because excluding silently
would be worse than the original bug: a board could pass by leaving the hard
pairs as stubs.

DRC continuity was tried as the gate first and is disqualifying — the
fixture is reverse-engineered, and continuity calls 25 of the human board's 49
pairs discontinuous, which would break the calibration anchor. The threshold
(0.5) is calibrated to the anchor's own thinnest leg, 0.522.

3. si_finish handed back a dirtier board than it was given. route_all
prunes as its last word, but si_finish runs after it and every stage rips and
re-routes copper: 231 dead traces and 52 dead vias left stranded and counted as
routed length. This one was invisible without the end-to-end run — the fix in
(1) was working inside route_all (2030 dead traces pruned vs 393 when
re-pruning the saved board) yet the final receipt came out byte-identical,
because everything cleaned up was replaced afterwards.

Results — fresh full route, 874.5s, routability 0.994

claim	before	after	bound
worst_group_skew	9.297mm	3.579mm	≤ 10.0	HOLDS
worst_intra_pair_skew	37.853mm	38.521mm	≤ 1.1	BROKEN
si_pairs_incomplete	(new)	9	≤ 0	BROKEN
min_pair_coupled_fraction	0.000	0.221	≥ 0.5	BROKEN
vias_per_si_net	2.766	3.106 → 2.626 w/ spur pass	≤ 3.0	BROKEN

Route-attributable Clearance +150 → +135, NetIslands 93 → 17.
MinDrill/AnnularRing/MinTraceWidth fire on nearly every via and thin trace
on the human board too, so they are imported-rule artifacts and Clearance is
the rule that attributes.

What is left

worst_intra_pair_skew is unchanged at 38.521mm. /MIPI1.D3_N holds 77.03mm of
copper for a 34.54mm span, continuous with 2/2 pads and all of it on
pad-to-pad paths — a redundant cycle, the signature of a re-route that never
removed the path it replaced. Leaf-pruning cannot touch it (every node has
degree ≥ 2), and compensation would add copper to a net that already has twice
what it needs.

Also staged but deliberately not wired in: spur_copper_mask /
prune_spur_copper, which strips dead-end branches off live islands
(/USB3-1.DP carried 72.86mm whose shortest pad-to-pad path is 36.60mm — a
~34mm unterminated stub on a USB3 pair). It is sound — it changes neither
board's UnconnectedNet count — and it recovers vias_per_si_net to 2.626, but
for a partially-reached net every piece of copper is a dead end, so enabling it
reclassifies those nets as unrouted and moves routability. That contract
change needs its own full-board before/after.

Guarantees

• No bound was changed. The human anchor re-verifies ALL HOLD after
  every commit and after the merge, with every pre-existing number identical
  (9.756 / 1.074 / 0.857 / 2.265).
• Stripped-fixture DRC baseline unchanged at 311 short/clearance, 23
  Clearance — the connectivity fixes do not move it.
• 299 tests pass, clippy -D warnings clean, cargo fmt clean. The new
  behaviour is pinned by dead_copper_shorted_to_a_live_net_is_still_pruned,
  verified to fail without the fix.
• No packages/kernel-wasm artifacts touched.

Full measurement record, including what did not work (descend_board tunes 0 of
41 pairs on a full board; phase compensation reaches 3 of 41; construction
couples 47 of 49 but negotiation un-couples down to 10 below 0.5), is in
docs/gpu-router-m7-pair-si.md.

🤖 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.

3 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
mecheval	Ignored		Jul 25, 2026 5:28pm
vcad	Ignored		Jul 25, 2026 5:28pm
vcad-docs	Ignored		Jul 25, 2026 5:28pm