Refine small cyclic layouts with pin-aware search

[pifuyuini]

The SI7021 reproduction from #11 leaves the two branches and their shared terminal scattered after greedy packing. This adds a bounded, pin-direction-aware refinement to small cyclic fallback partitions, placing the branches side by side and centering the terminal beyond them.

Related to #12. This is a general search alternative to the topology-specific proposals there; it does not claim a lower SI7021 Manhattan length than their reported 3.9.

SI7021 placement metric	Original packer	This change
Strong pin-to-pin Manhattan length	5.9	3.9
Length estimate including outward pin stubs	7.05	4.1
Longest pin-to-pin connection	2.175	0.95
Actual routed length of the same five strong connections	6.7	3.9

The pass applies only to cyclic fallback partitions with 3–6 chips. It preserves the anchor, fixed placements and non-cycle branches, respects allowed rotations and pin/body clearance, and rejects added straight-line crossings or third-component obstructions. A fixed-seed search escapes greedy placement traps; connected free groups then move together and receive local alignment refinement. Equal-length results favor balanced branches. Specialized layout strategies bypass this pass.

Rotation geometry is cached. Each acceptance step consumes its random draw consistently so rounding-sized input differences do not alter the random sequence. The included bun benchmarks/refine-cyclic-partition.ts runs five warmups and twenty timed samples. On the same local machine, this pass's median fell from 85.35 ms before caching/refinement to 30.14 ms; this is a single-case measurement, not a cross-machine performance guarantee.

Before	After

Validation: 125 tests pass, 1 pre-existing skip, 0 failures. bun run build, tsc --noEmit, formatting and whitespace checks pass. Eleven focused tests cover the original fixture through the full pipeline, clearance, fixed positions/rotations, deterministic results, renamed components, rounding-sized coordinate differences, a quarter-turn input, synthetic three/six-chip rings, and a shorter-but-obstructed candidate that must be rejected. SVG output was rendered and inspected. The synthetic rings demonstrate different graph shapes, not additional real-world circuits.

The new downstream integration test runs the original and refined placements through @tscircuit/schematic-trace-solver 0.0.69, using the 2.4 maximum pair distance from @tscircuit/core 0.0.1331 and allowing all net-label directions. It verifies physical endpoints for all five strong connections, orthogonal final paths, no component-interior intersections, and preserved named-net labels after refinement. The five actual routes total 6.7 before and 3.9 after; this comparison excludes other label/net routing choices.

Limitations: this is a downstream trace-pipeline integration test, not full @tscircuit/core circuit rendering. Both layouts have one unsuccessful GND routing attempt and use matching GND labels; the test explicitly verifies that fallback rather than treating solver.solved as proof that every attempted pair has a physical route. Results remain specific to this fixture and configuration. The fixed search budget adds work to qualifying small partitions.

Developed with AI assistance. Legacy bounty clarification: docs #480 moved the current program to contributor sponsorships. Does the historical $150 offer in #12 still apply to an accepted solution, with PayPal payout, or is this work eligible only for the current GitHub Sponsors program? No payment or eligibility is assumed from the old marker.

/claim #12

[vercel]

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

Project	Deployment	Actions	Updated
matchpack	Ready	Preview	Sep 20, 2026 6:33pm UTC

[pifuyuini]

Added a downstream routing integration test: with schematic-trace-solver 0.0.69 and core's 2.4 distance threshold, the same five strong connections route successfully and their actual path length falls from 6.7 to 3.9. It checks endpoints, orthogonality, component clearance, and named-net labels. Both layouts use GND label fallback; this is not a claim that every attempted pair has a physical route. Local suite: 125 pass, 1 existing skip; build/typecheck pass.

One payment clarification before further work: I found that docs #480 switched the program to sponsorships. Does the historical $150 offer in #12 still apply if this solution is accepted, and can it be paid through PayPal, or is only the current GitHub Sponsors program available? The implementation and validation are AI-assisted, as disclosed in the PR.