Preserve hard constraints in greedy final routing

[ShiboSoftwareDev]

Stack

Stacked on reproduction PR #139.

Problem

Normal path search uses Infinity to reject hard-invalid moves such as crossing another net in a single-layer region.

Two shortcuts combined:

1. the timeout-only greedy solver returned the current path cost directly, which removed both soft costs and hard rejections;
2. the base path loop accepted the final hop into an endpoint before calling the cost/legality calculation.

Fix

The fix has two matching steps:

1. every final hop now uses the same legality calculation as every intermediate hop;
2. greedy routing keeps a non-finite hard rejection, but still discards finite soft costs.

The shared final-hop check also replaces the duplicate implementation that previously existed only in the distance-aware subclass. This does not add a new stage, topology type, or downstream geometry patch.

Reproduction and visualization

The unchanged #139 fixture has a vertical net already routed through one single-layer region. Greedy completion tries to add a horizontal net whose only path crosses it.

Before: reproduction PR #139	After: this fix

The SVG is tiny-hypergraph’s native solver visualization:

• a solid line is a route accepted into solver state;
• a dashed line is an uncommitted route request.

Before the fix, both crossing lines are solid and the center region reports one same-layer crossing. After the fix, only the legal vertical route is solid; the horizontal request stays dashed and the center reports zero crossings.

Validation

• GitHub-hosted full test and snapshot suite passed
• autorouter sample 4 integration benchmark: pending after the upstream SHA is pinned
• nothing is run locally

[vercel]

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

Project	Deployment	Actions	Updated (UTC)
tiny-hypergraph	Ready	Preview	Aug 2, 2026 10:51am

[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.
The result comment shows the matching dataset from the latest stored main-branch benchmark artifact beside the PR run; main is not rerun for each PR benchmark request.

Examples:

• /benchmark -> hg07, all samples, core solver
• /benchmark --dataset srj18 --limit 20 -> first 20 srj18 samples
• /benchmark --dataset srj18 --concurrency auto -> srj18 samples with runner CPU-count concurrency
• /benchmark --sample sample002 -> one hg07 sample
• /benchmark --limit 20 --solver poly -> first 20 hg07 samples with the poly solver
• /benchmark --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.