Wide-spacing queries on layers without wide parts

A layer can declare a wide-spacing rule while containing only narrow metal. The wide-spacing owner index contains only actual wide parts, so such a layer has no reach entry. A query on that layer has no same-layer owner candidates. It must consume the empty-query work unit and return from that visit, rather than indexing a missing entry and panicking. Queries on other layers, same-net notch checks, and active-layer comparison counts stay unchanged.

The mixed-layer regression first reproduces the original panic: M1 has a real wide-spacing violation, while M2 has a rule but only narrow wires, via pads, patches and a pin. It then adds a wide M2 component and checks both distinct-net and same-net violations. A low comparison cap must still fail explicitly.

Upfront resource estimate

The fixture has four wires, at most three patches, three pins and one via on two layers; its bounded rectangle/grid state is conservatively below 1 MiB per test and adds no path enumeration. Existing engine comparison, grid and piece caps remain in force. The new guard allocates nothing.

The fresh two-worker build was estimated below 3 GiB aggregate from prior measurements. All checks use a 4 GiB RAM / 1 GiB swap cgroup, bounded wall time, CARGO_BUILD_JOBS=2, RUST_TEST_THREADS=2, CPUs 24–31, and qualified GNU time. The serial dual-PDK harness estimates below 1 GiB per native child and uses the same route budgets, pinned historical executable and independent DEF geometry comparison as the existing SKY130/GF180 controls. No high-memory scale lock is needed for these estimates below 4 GiB.

The C5 diagnostic replay retains its original admission estimate of 3955 MiB plus a 32 MiB controller, 16 requested route threads, 1200-second route timeout, and 4 GiB / 1 GiB cgroup. It runs only in a newly copied private rev3 workspace. Original failure evidence and qualified binaries remain immutable. Timings may overlap GPU compilation and cannot establish controlled speedup.

Validation

The focused regression failed at the original missing-key panic, then passed after the guard. Workspace tests passed: 435 passed, 0 failed, 11 existing ignored; fresh build plus tests took 91.26 seconds, peak RSS 972084 KiB. Formatting, Clippy on all targets/features with warnings denied, and rustdoc passed. Raw resource and command receipts are private under .work/perf-20261030/qor-c5-fix/ in the superproject.

The release build passed in 102.01 seconds with 1011632 KiB peak RSS. The unmodified dual-PDK harness passed all 36 CLI commands, exact routes/netlists, and six independently expanded DEF geometry comparisons against the pinned historical executable. Both SKY130 and GF180 inputs were hash-rechecked. The fixed private executable SHA-256 is 2917b58c71006e40170f5aceeedb16abbae4666c516afa6417167e01d6d3afdb. This validates the focused guard and existing routing controls; IHP C5 QoR and foundry DRC/LVS/timing remain separate diagnostic gates.