plan-pcb-placement

Plan PCB Placement

Placement decides what routing can achieve, and a wrong placement cannot be rescued by any router. This skill owns that half: deciding whether to touch the placement at all, fixing it when it is wrong, and proving the fix.

It never routes. When the board is placed and you want copper, use /plan-pcb-routing. When one run must do both, use /plan-pcb-placement-and-routing, which sequences the two and owns the rules that only exist when they meet.

What you are optimising FOR

A board that ROUTES: parts arranged so they work together, ending at zero unrouted and zero broken. Not parts returned to the poses they used to have. On a repair or reconstruction job this distinction decides which result you call a success.

• Lead with the routed outcome. board_score's blocking — unrouted + broken first — then quality as the tie-break once it is 0.
• On the perturbed corpus, recovery and home /N are DIAGNOSTICS, not the score. They measure distance to the original pose. Measured (run 10): recovery −0.0014, home 0/30, on a run that took copper-free DRC 9 → 0, assembly blocking 4 → 0, and the board from NOT BUILDABLE to buildable. The one recovery figure that IS a defect signal is collateral_pad_rms rising — that means parts nothing had damaged were moved.
• A human original is a benchmark to approach (its vias / copper / segment counts), not a pose to match.

The top-priority defect: pad copper off the outline

Ahead of every clearance graze. Those nets cannot be routed at all, so the defect converts one-for-one into unrouted and broken. Measured, run 10: 11 such parts caused ALL 13 unrouted nets and most of 37 broken ones.

Read it from render_placement --json-out's checklist.a_off_outline.pad_copper. A whole-board pass/fail verdict is the wrong channel to learn it from — check the per-part list.

Scope the search to the refs the gate names

When a gate names specific parts, free exactly those and lock everything else. A whole-board sweep orders its violators by its own priority — usually worst-off-board first — and may never reach the ones blocking you. Measured: 2 parts freed and 105 locked cleared both blocking pairs in 63 seconds, where whole-board sweeps ran 10+ minutes without touching them.