Adds pcbComb, a <trace> prop in the same family as the existing pcbStraightLine and pcbPath — a fixed, manually-shaped route rather than an autorouted one. Where pcbStraightLine draws a direct line, pcbComb draws the other common hand-routed shape between two lines of pads: a perpendicular escape off the pad, one 45° diagonal, and a perpendicular landing. N of them nest into an even comb.

<trace from=".U1 > .p1" to=".U2 > .q8" pcbComb="rowToColumn" />

The value is <sourceLine>To<targetLine>: columnToColumn | rowToColumn | columnToRow | rowToRow (a COLUMN of pads escapes perpendicular in x, a ROW in y). The two bend points are derived from the pads' live global positions in the manual-trace phase, before autorouting — so the comb re-derives if a part moves, nothing frozen. A comb whose fixed shape can't reach its target without backtracking is left unrouted for the autorouter.

Where it helps — and where we're unsure

We hit this on a denser real board and wanted to offer a distilled, deliberately contrived example rather than imply it's common. The same 12-net row→column bundle, routed three ways on the identical board:

autorouter

pcbStraightLine

pcbComb

To be clear about what this isn't: the autorouter isn't wrong — it optimizes each net with no signal to keep a bundle uniform, and just gets erratic under congestion (a 10 mm change in the gap flips it between 0 and 14 vias). For a plain fan-out it ties pcbComb; the corner turn is where a declared shape earns its keep. And we honestly don't know whether this case is common, whether we're doing something else that creates our need, or whether bundle-shaping is a direction tscircuit wants to own vs. leave to the autorouter. We're offering it, not arguing for it.

The change

• lib/utils/computeCombWaypoints.ts — pure geometry (the two bend points).
• A branch in Trace_doInitialPcbManualTraceRender mirroring the existing pcbStraightLine branch, plus a matching skip in Trace_doInitialPcbTraceRender. No autorouter changes.
• A snapshot test for the comb. docs/pcb-comb/ is only the illustration above — happy to drop it.

On the red CI: both the type-check and test failures here are only the props dependency — <trace pcbComb="…"> isn't in the upstream traceProps yet (the type and the runtime zod schema), so type-check rejects the prop and the test's trace autoroutes instead of combing. Both clear once tscircuit/props#715 publishes; nothing else is failing (the other test shards are just fail-fast cancellations).