feat: add physical-capacity routing with HighDensitySolverB02

[seveibar]

What this does

Bugreport101 looks like a 37-connection board, but expands to 190 terminals, 155 post-escape point-pair routes, 718 physical-capacity high-density nodes, and 1,173 exact local route pairs. The original cache-free Pipeline 7 effort-1 solve took about 395.9 seconds.

This PR replaces the earlier search-budget heuristic with a new routing path:

• physical-capacity portal pathing uses only authoritative legal shared-edge slots and never fabricates sub-pitch duplicate terminals;
• HighDensitySolverB02, supplied by the pinned @tscircuit/high-density-b01 package, composes a bounded B01 partial-route seed with conflict-directed planar and cross-layer repair, then publishes output only after strict trace/via clearance validation;
• a thin autorouter adapter preserves explicit pair topology and order, terminal PCB port IDs, root connection names, and region IDs while retaining the existing duplicated-portal and node-obstacle applicability gates;
• TwoChordLaneIntraNodeSolver handles the common perpendicular two-chord channel case deterministically;
• every specialized winner is accepted only after endpoint/root/region preservation, via-transition and bounds checks, package geometry validation, and 0.10 mm copper-clearance validation.

The exact former hotspot topology_merge_19__sub_0_2 fell from about 51.6 seconds to about 0.15 seconds. Final isolated high-density routing fell from 105.3 seconds to 13.1 seconds, with 63 HighDensitySolverB02 winners and 4 two-chord winners.

End-to-end result

Two clean opt-in runs on the reference machine:

Metric	Previous optimized draft	New solver path
Total solve	116.1 s	43.7 s / 44.3 s
Port-point pathing	8.8 s	7.5 s
High-density routing	51.2 s	13.1 s / 13.6 s
Exact DRC improvement	12.8 s	6.6 s / 6.8 s
Power expansion	21.7 s	9.6 s
Relaxed DRC errors	47	22

The post-integration run using merged HighDensitySolverB02 completed in 41.808 seconds with the same 2,498,457 iterations and 22 relaxed DRC errors. The original <120_000 ms contract remains unchanged. This is not a claim that the whole board is DRC-clean; the remaining 22 issues stay visible and regression-guarded.

Scope and safety

The output-changing route is explicit and default-off via:

highDensityRoutingMode: "physical-capacity"

The Bugreport101 test and Cosmos fixture opt in. Legacy Pipeline 7 callers retain duplicate-port preprocessing, uniform redistribution, one-step pathing scheduler granularity, and the existing HD portfolio. The previous initial/grown iteration caps, largest-scale-first ordering, and unvalidated post-shrink repair-seed policy have been removed from this branch.

Supporting optimizations include:

• batched tiny-hypergraph parent stepping in physical-capacity mode, with strict option validation;
• exact-geometry evaluator memoization;
• per-route spatial-conflict memoization during useless-via removal;
• reusable immutable obstacle and same-net metadata in power expansion.

The GitHub-only Cosmos verification step uses a 6 GiB Node heap. Its 4.1 GiB default heap consistently aborted while rendering the 4,881-module fixture bundle; the external Vercel deployment was already green.

Visual snapshot

The env-gated regression includes this current routed snapshot in addition to the always-reviewed unrouted input snapshot:

Upstream dependencies

• feat: add conflict-directed HighDensitySolverB02 high-density-b01#9 — numbered HighDensitySolverB02, composing bounded B01 seeding with strict conflict-directed repair; pinned to merge commit f974f97a77f27bc32b1fe5a8b3bccb7ba023e0b7
• perf: cache canonical connection names power-trace-expander#24 — canonical connection-name cache
• perf: reuse static spatial obstacle metadata power-trace-expander#25 — reusable static spatial metadata; exact power phase 38.66 s to 9.09 s with byte-identical output
• perf: cache exact DRC connectivity and consume HD routes directly high-density-repair03#86 — connectivity and geometry-evaluator caches
• Optimize dense candidate pathing hot loop tiny-hypergraph#174 — dense candidate fast path

All dependencies are pinned by commit SHA.

Validation

• env-gated full Bugreport101 solve after B02 integration: 41,808 ms, 2,498,457 iterations, and 22 relaxed DRC errors
• routed SVG snapshot matched without update mode
• focused HighDensitySolverB02 integration coverage: 2 tests, 2,754 assertions, 0 failures
• focused coverage verifies the upstream-backed portfolio winner, deterministic dominant-node and cross-layer output, exact repeated-name pair identity, non-null terminal metadata, bounds, vias, obstacle gating, and strict geometry clearance
• related physical-capacity, two-chord, and legacy-isolation tests pass
• bunx tsc --noEmit --pretty false passes
• bun run build passes
• git diff --check passes

Run the full regression with:

RUN_BUGREPORT101_CM5_SPI_ROUTING_TIMEOUT=1 bun test tests/bugs/bugreport101-cm5-spi-routing-timeout.test.ts --timeout 9999999

The runtime assertion remains env-gated because wall time is hardware-sensitive. The deterministic exact-node tests use iteration, route-identity, and geometry assertions instead of wall-clock thresholds.

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
capacity-node-autorouter	Ready	Preview	Aug 29, 2026 9:34pm

[tscircuitbot]

Benchmark This PR

Run benchmarks by commenting on this PR:

/benchmark [benchmark.sh args...]
/benchmark-long [benchmark.sh args...]
/benchmark-all [benchmark.sh args except --dataset]
/benchmark --same-machine
/benchmark-all --same-machine
/profile --dataset 18

Comment /benchmark to run the default dataset, or append any arguments accepted by ./benchmark.sh.
Comment /benchmark-long for an 8-vCPU run that defaults to 8 workers and has an eight-hour timeout.
Comment /benchmark-all to start separate workflow runs and result comments for the default dataset plus srj18, srj19, srj20, srj21, and srj23. You may append benchmark arguments such as --pipeline 9; --dataset is not accepted because the command selects every configured dataset.
Append --same-machine to /benchmark or /benchmark-all to compare current main and the PR head sequentially on the same 8-vCPU Blacksmith runner.
Comment /profile --dataset 18 to compare direct Pipeline 7 stage-time percentages between current main and the PR head sequentially on one Blacksmith runner.

Everything after /benchmark, /benchmark-long, or /benchmark-all is safely forwarded to ./benchmark.sh, except --profile-solvers, which enables profile comparison tables, and --same-machine, which enables the paired runner.
Examples: /benchmark --dataset 18 --sample-timeout 2000s, /benchmark --pipeline 10 --dataset 29, /benchmark --pipeline 10 --dataset 29 --sample 1, /benchmark --pipeline 7 --limit 20, /benchmark all 20 --concurrency 8, /benchmark --same-machine, /benchmark-all --pipeline 9 --same-machine, and /benchmark-long --dataset 18.

Use /update-snapshots (or /us) to run BUN_UPDATE_SNAPSHOTS=1 bun test --timeout 120_000 on the PR branch and auto-commit snapshot updates.
Use /usf to read recent failed test files, update and verify their exact CI test shards, and auto-commit only their snapshots. It uses the configured fast benchmark runner by default; use /usf --ubuntu-latest for GitHub-hosted x64 CI parity.

Any PR whose title contains [BENCHMARK TEST] will automatically run one default-dataset benchmark on PR updates; it does not post a PR result comment.

[seveibar]

/benchmark --same-machine --dataset 18 --pipeline 9

[seveibar]

/benchmark --same-machine --dataset 18 --pipeline 7

[tscircuitbot]

Same Machine Benchmark Results

Both revisions ran sequentially in one Blacksmith job on blacksmith-01m15ak03fkehhpndpd7cz19kv-8vcpu.

Dataset: srj18 · Scenarios: 16
Main: 96645b4 · PR: 343bc51

Solver	Metric	Main	PR	Delta
Pipeline7	Completion	81.3% (🕒6.3%)	81.3% (🕒6.3%)	0.0 pp
Pipeline7	Relaxed DRC pass	43.8% (🕒6.3%)	43.8% (🕒6.3%)	0.0 pp
Pipeline7	DRC issues	125	125	0
Pipeline7	Timeouts	1	1	0
Pipeline7	P50 time	95.9s	96.0s	+0.0%
Pipeline7	P60 time	104.5s	113.1s	+8.3%
Pipeline7	P70 time	165.0s	154.9s	-6.1%
Pipeline7	P80 time	225.8s	233.8s	+3.6%
Pipeline7	P90 time	291.1s	278.1s	-4.4%
Pipeline7	P95 time	329.3s	311.9s	-5.3%
Pipeline7	Average vias	210.62	210.62	0.0%

Outcome changes: 0 improved, 0 regressed. DRC issues are totaled across solved samples. Timing percentiles include solved and timed-out samples; negative timing deltas are faster.

Workflow: View run
Artifact: Download raw reports

[tscircuitbot]

Same Machine Benchmark Results

Both revisions ran sequentially in one Blacksmith job on blacksmith-01m15ak04kwbdw3j30aqytrcdy-8vcpu.

Dataset: srj18 · Scenarios: 16
Main: 96645b4 · PR: 343bc51

Solver	Metric	Main	PR	Delta
Pipeline9	Completion	75.0% (🕒12.5%)	75.0% (🕒12.5%)	0.0 pp
Pipeline9	Relaxed DRC pass	50.0% (🕒12.5%)	56.3% (🕒12.5%)	+6.3 pp
Pipeline9	DRC issues	74	75	+1
Pipeline9	Timeouts	2	2	0
Pipeline9	P50 time	103.7s	107.7s	+3.9%
Pipeline9	P60 time	132.8s	130.5s	-1.8%
Pipeline9	P70 time	230.8s	197.9s	-14.2%
Pipeline9	P80 time	274.5s	261.2s	-4.9%
Pipeline9	P90 time	352.6s	350.6s	-0.6%
Pipeline9	P95 time	360.0s	360.0s	0.0%
Pipeline9	Average vias	187.50	186.17	-0.7%

Outcome changes: 1 improved, 0 regressed. DRC issues are totaled across solved samples. Timing percentiles include solved and timed-out samples; negative timing deltas are faster.

Changed outcomes (1)

Solver	Sample	Main	PR	Main time	PR time	Delta
Pipeline9	8	Solved (DRC failed)	DRC passed	234.1s	195.9s	Improved

Workflow: View run
Artifact: Download raw reports

[seveibar]

Superseded by focused PRs:

• superseded: available segment-point capacity #2296 physical portal-capacity correctness
• feat: add automatic TwoChordLaneIntraNodeSolver #2297 TwoChordLaneIntraNodeSolver
• feat: add automatic HighDensitySolverB02 adapter #2298 HighDensitySolverB02 adapter
• perf: cache routing geometry evaluations #2299 reusable routing performance work
• feat: route Bug 101 with explicit capacity and batching #2300 Bug 101 Pipeline 7 rollout
• superseded: Pipeline 9 B02 coverage moved to #2298 #2301 Pipeline 9 B02 rollout
• feat: add opt-in GrowShrink original-scale validation #2302 GrowShrink fail-loud correctness follow-up (draft)

The first five are a clean dependency stack. Pipeline 9 branches from the B02 adapter, while GrowShrink remains independent because Bug 101 currently exposes nodes that still rely on the legacy invalid fallback.