perf(router): negotiation-first ordering cuts the CM5 route 113min → 14min

[ecto]

What

The greedy phase paid for its unrouted tail three times over.

1. Every connection the speculative batch search missed got the full sequential rescue arsenal (flow escape, coupled-pair phantom, true shove) inline in the batch loop, one connection at a time.
2. The GPU negotiator — which resolves that same tail in seconds — only ran afterwards.
3. Then rip-up rounds re-solved it a third time, and the incremental negotiation rounds a fourth.

Measured on the full CM5, the baseline spends 82% of its 113 minutes on three rounds that all end below where they started and are then discarded wholesale by keep-best:

baseline phase	cost	outcome
route_batch (greedy + inline arsenal)	680.5s	500 conns → 37 unrouted
GPU negotiation	~10s	37 → 8 pending
rip-up round	2821.6s	placed 536 → 511, discarded
incremental round 1	1218.4s	placed 536 → 494, discarded
incremental round 2	1506.5s	placed 536 → 475, discarded
post-loop repair / stitching	~550s	—

This reorders the layer-rich regime around the negotiator and lets the arsenal work only what negotiation leaves behind.

Changes

• route_batch runs search-only when negotiation goes first. Failures fall straight through to the negotiator; the arsenal fires once afterwards over its leftovers (new arsenal_sweep). On the full CM5 negotiation takes pending 37 → 11 in 4.9s, the sweep 11 → 10 in 48s. Dropping the inline arsenal also cuts the greedy batch itself from 680.5s to 470.9s for identical output (191 committed, 37 unrouted).
• Skip the greedy phase's rip-up rounds once negotiation has run first — that 47-minute round above, plus 8 minutes on the 250-net subset (382 → 380), both discarded.
• Stop the negotiation loop after round 0 in this regime. Round 0 is already a negotiated pass carrying the arsenal and the last-round fine-retry that the trailing polish round exists to guarantee, so stopping forfeits nothing it could have added.
• Make the surviving budgets binding. They were checked between connections only, but one connection's victim re-route is itself a whole batch — so a 92s budget measured 231s. route_batch now takes an optional deadline; past it the remaining queue is reported unrouted, which every caller already handles (a stuck connection stays pending, a ripped victim gets its original copper restored). It costs attempts, never legality.

Board-regime rule preserved

The >= 4 copper layers gate is unchanged and now lives in one place (negotiate_first_regime). Thin boards keep arsenal-first ordering, the full rip-up ladder, and all four negotiation rounds — their scarce corridors are spent more carefully by the arsenal (the VESC lost 4 points of routability to negotiation-first). Boards built without the gpu feature, or run without VCAD_GPU_NEGOTIATE=1, are untouched.

Benchmarks

Same machine, same session, background load ~5, VCAD_GPU_BATCH=1 VCAD_GPU_NEGOTIATE=1. CM5 before/after are both measured here, not quoted:

board	before	after
CM5 (10 layers, 522 conns)	6781.4s = 113.0 min @ 0.985	863.9s = 14.4 min @ 0.981	7.85× faster
VESC (thin board)	~26s @ 0.79	22.9s @ 0.787	unchanged
moteus (Eagle)	80–124s @ 0.87	107.6s @ 0.891	unchanged / slightly up

Where the CM5's 14.4 min goes now: route_batch 470.9s, GPU negotiation 4.9s, arsenal sweep 48.4s, post-loop repair and plane stitching the rest.

Routability trade: 0.985 → 0.981, i.e. 2 connections of 544, against 99 minutes. It is above the 0.98 bar but worth naming: in the old order the inline arsenal commits its rescues during the batch, so the board those rescues leave behind lets negotiation reach pending 8 where the new order reaches 10. Recovering them would mean running the sweep over all 37 failures before negotiation rather than over its 11 leftovers — roughly +2 minutes — if those two connections are worth it.

Verification

• cargo test -p vcad-ecad-pcb — 248 passed
• cargo clippy -p vcad-ecad-pcb --features gpu -- -D warnings — clean (also clean without the feature)

🤖 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 12:58pm
vcad	Ignored		Jul 25, 2026 12:58pm
vcad-docs	Ignored		Jul 25, 2026 12:58pm
vcad-mcp	Ignored		Jul 25, 2026 12:58pm