[JeffMboya]

What type of PR is this?

This is a bug fix.

What does this do?

island_vias now walks track, pad and via chains through join_through_copper, so a pad-holding sliver is not an island.

Which issue(s) does this PR fix/relate to?

Fixes #367.

Have you included tests for your changes?

Yes. A track through the gap, a 0.01 mm gap staying apart, and overlapping pads joining.

Did you document any new/modified features?

No, an internal repair; no user-facing change.

Notes

Copper joins only where it touches (1 nm tolerance, was 20 um); overlapping same-net pads join. fmt, clippy clean.

[SammyOina]

Changes requested

[P1] Detect track copper intersecting a pour, not only track endpoints

pieces_of() associates a track with a pour only when one of its endpoints is inside the pour region:

let mut regions = hits(&segment.layer, a);
regions.extend(hits(&segment.layer, b));

A track that passes through a sliver or pour with both endpoints outside it is still electrically connected in KiCad, but this implementation will not associate it with that region and may incorrectly add a redundant island via.

The new chain logic also treats tracks as zero-width and uses only centerline/end-point proximity. It can miss valid connections where the track copper overlaps a pad, via, or another track by its width.

Please use segment/polygon intersection and copper width when determining connectivity. Add regression coverage for:

• a track crossing a pour with both endpoints outside the pour;
• a track ending at the edge of a pad/via using its actual width;
• a via chain connecting the sliver to the main pour.

The current test only covers two centerline-aligned tracks whose endpoints meet, so it does not exercise these physical-copper cases. CI is green, but the connectivity model is still narrower than KiCad’s.

[JeffMboya]

@SammyOina Fixed in 8b9050b: segment/pour intersection and width-aware track, via and pad joins; four regression tests added.

[SammyOina]

Changes requested\n\n### [P1] Detect pad and via copper overlapping a pour edge\n\nIn crates/synth-router/src/island_vias.rs, pieces_of associates pads and vias with pours only when their centers are inside the region. A large pad or via whose copper overlaps a pour edge while its center is outside will therefore remain disconnected in the model, and the repair can still add a redundant island via.\n\nUse the actual pad and via copper geometry against the region, not center-point containment, and add regression fixtures for a pad-edge overlap and a via-edge overlap.\n\n### [P2] Avoid sampled track-to-pad connectivity\n\ncopper_gap samples track-to-pad distance every 0.05 mm. A short overlap between samples can be missed, particularly for narrow pads or angled tracks. Replace the sampling with exact segment-to-pad geometry, or use a conservative bound that cannot miss an overlap, and add a fixture where the overlap is shorter than the sampling interval.\n\n### [P2] Bound the connectivity pass for large nets\n\njoin_through_copper compares every pair of same-net pads, vias, and tracks. Large ground nets can make this quadratic, and track-to-pad sampling adds further work. Add a realistic large-board performance test and use spatial bucketing or another spatial index if the pass is not bounded sufficiently.\n\nThe earlier track-through-pour and width-aware track/via-chain concerns are addressed, but the pad/via-to-pour center test leaves the physical-copper connectivity model incomplete.

[JeffMboya]

@SammyOina 5af8d04: pads and vias join pours by their copper, track-to-pad distance is exact, the pass is grid-bounded; edge, corner and 12,000-piece fixtures added.

[SammyOina]

Changes requested

[P1] Do not treat a 20 µm real gap as electrical connectivity. join_through_copper currently unions pieces when copper_gap <= 0.02. This means two same-net tracks, pads, or vias separated by a genuine 0.01 mm gap can be classified as connected, suppressing the island via that is needed to restore connectivity. Use exact geometry or a much smaller numerical epsilon, and add a regression proving that a small but nonzero gap remains disconnected while true copper overlap remains connected.

[P2] Handle overlapping same-net pads. copper_gap returns f64::MAX for Pad–Pad pairs, so overlapping pads are never joined through direct copper. KiCad can use overlapping same-net pads for stacked, thermal, or exposed-pad footprints. Add direct pad-overlap handling and a regression if those footprint types are supported by the exporter.

The previous track-crossing, copper-width, pad/via edge-overlap, exact track-to-pad, and large-net performance concerns are addressed, and CI is green. Please resolve these remaining connectivity cases before approval.

[JeffMboya]

@SammyOina 3a65744: copper joins only on touch (was 20 um), overlapping same-net pads join; regressions for both.

[SammyOina]

Changes requested

[P1] Preserve the actual outline of non-rectangular pads in connectivity

PadBox::new() reduces every pad to a sharp rectangle before pieces_of() and join_through_copper() decide whether copper is connected. That is wider than the actual copper at the corners for KiCad circle, oval, and roundrect pads. A track, via, or pour that falls only in one of those rectangular corner areas can therefore be classified as connected even though there is a real clearance gap. The island-repair pass would then suppress the tie via that restores the physical connection.

The PCB reader already retains Pad.shape and corner_ratio, and its signed-distance implementation distinguishes these outlines. Please carry shape-aware geometry into this connectivity path (or use the shared geometry), rather than treating every pad as a rectangle. Add regressions for circle, oval, and roundrect corner-clearance cases that stay disconnected, alongside a touching case that joins.

The previous track-through-pour, physical-width, edge-overlap, small-gap, pad-overlap, and large-net cases are addressed by the current head; this is the remaining blocker for approval.

[JeffMboya]

@SammyOina b36ccef: pads join by real outline (circle, oval, roundrect radius); corner-clearance regressions stay apart, touching joins.