Preserve downstream routing improvements

[seveibar]

Summary

• require downstream pressure metrics to improve monotonically during region-cost optimization
• prevent path relinking from reintroducing layer transitions
• keep region-cost improvements that make the downstream high-density search easier

Validation

• 144 tiny-hypergraph tests pass
• TypeScript check passes
• srj18 sample 2: 0 DRC errors, 787,630 high-density iterations
• previous broad candidate: 4 DRC errors, 883,527 high-density iterations

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
tiny-hypergraph	Ready	Preview	Aug 19, 2026 11:35pm

[tscircuitbot]

Benchmark This PR

Run benchmarks by commenting on this PR:

/benchmark [benchmark.sh args...]
/benchmark --dataset <hg07|srj18> --limit <n> --solver <core|poly> --families <default|default+deep|all> --concurrency <n|auto>

Everything after /benchmark is forwarded directly to ./benchmark.sh.
Every command runs current main and the PR head sequentially on the same Blacksmith machine, then reports their results side by side.

Examples:

• /benchmark -> all committed completed SRJ18 Pipeline7 cases, core solver
• /benchmark --limit 4 -> first four committed Pipeline7 cases
• /benchmark --dataset srj18 --concurrency auto -> srj18 samples with runner CPU-count concurrency
• /benchmark --sample sample003 -> one committed SRJ18 Pipeline7 case
• /benchmark --dataset hg07 --limit 20 --solver poly -> first 20 hg07 samples with the poly solver
• /benchmark --dataset hg07 --limit 40 --families default+deep -> first 40 hg07 samples with deeper candidate families

Any PR whose title contains [BENCHMARK TEST] will automatically run the benchmark workflow on PR updates.

[seveibar]

Dataset-18 validation found this is useful locally but not safe globally.

• srj18 sample 2: 0 DRC errors, 787,630 high-density iterations; repaired prior broad candidate had 4 errors / 883,527 iterations.
• Same-machine full dataset: DRC pass improved vs autorouter main from 43.8% to 50.0%, but completion stayed 87.5% and P95 rose 468.6s -> 2537.1s.
• Against the immediately previous optimizer PR version, DRC pass stayed 50% while completion fell 93.8% -> 87.5%.
• sample 6 regressed from solved in 906s (85 DRC issues) to the full 7,200s timeout.

Conclusion: monotonic aggregate downstream pressure is not a sufficient proxy for detailed-router branching. Do not merge this as-is. Benchmark: https://github.com/tscircuit/tscircuit-autorouter/actions/runs/32314092070