ecad: HDI - back-side placement, microvias, via-in-pad, pinned copper by ecto · Pull Request #927 · ecto/vcad

ecto · 2026-10-15T12:26:02Z

What the rose brain board (RK3588S2, 0.4 mm FCCSP1253) needs from vcad.

Back-side placement: [EcadPlaceBack ref x y rot]. KiCad's flip: front/back layers swap, local y mirrors, pad angles negate. It runs on both the IR the router sees and the library body the writer emits. A mixed-side test board routes with 0 KiCad DRC violations, including no lib_footprint_mismatch.
Hand copper: [EcadVia], [EcadViaSpan x y d drill net from to] and [EcadTrace] are now converted. Before this, the types existed but convert ignored them. They are pinned as CopperSource::Manual, and the second routing pass and leaf trimming never rip them.
Microvias: vcad route --microvia D,DRILL treats outer-to-adjacent spans as microvias. The maze clears each via at its span's size, the CLI writes them at that size, and the writer emits (via micro …).
Via-in-pad: a pinned via inside a same-net pad makes that pad reachable on the via's layers. A BGA ball then leaves on the inner layer instead of the router drilling a via between balls.

Test: a 0.5 mm BGA-153 with a microvia in each signal ball gets 10 of 11 nets out, against 7 of 11 with through vias only.

Since then (brain board work)

Speed / visibility: --fast; stderr progress per phase; OpenTelemetry traces (OTEL_EXPORTER_OTLP_ENDPOINT); per-phase checkpoints (<out>.ckpt/, VCAD_ROUTE_REDO, VCAD_ROUTE_CHECKPOINT=0)
Repairs no longer hang: frozen zone fills, net-scoped continuity, one session per net, only stranded pads stitched, 10 min budget per repair; rip-up victim cap and early stop; second pass is one greedy round
Bug fixes:
- HDI via stacks (F-In1 + In1-In3 at one point) lost their lower hop in the maze, dedupe and leaf trim — the single biggest win on the brain board (490 → 316 unconnected)
- the ratsnest skipped every net that already had a trace; it now only skips pads joined by copper
- 10-layer boards: EcadBoard 10, In7/In8 in the stackup and in the connectivity layer stack
New: --only NET, --escape LAYERS (multi-layer BGA escape planner, off by default — not a win yet), VCAD_ROUTE_SKIP=links,stitch

🤖 Generated with Claude Code

…way KiCad flips PcbLayer::flipped; kicad_board::flip_to_back mirrors local y, negates pad angles and swaps front/back layers on the IR the router sees; the writer emits the library body the same way (layers, coordinates, mirrored text) so KiCad loads it with no lib_footprint_mismatch. A mixed-side test board routes with 0 DRC violations. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

- loon: [EcadVia ...], [EcadViaSpan x y d drill net from to] and [EcadTrace ...] lay copper by hand, pinned (CopperSource::Manual). - route --microvia D,DRILL: outer-to-adjacent spans are microvias; the maze clears them at that size and the CLI writes them at it; KiCad gets them as (via micro). - via-in-pad: a pinned via inside a same-net pad makes the pad reachable on the via's layers, so a BGA ball leaves on the inner layer instead of the router drilling between balls. - the second pass and leaf trimming never rip hand-laid copper. A 0.5 mm BGA-153 with a microvia in each signal ball: 10/11 nets out (was 7/11 with through vias only). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

vercel · 2026-10-15T12:26:06Z

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

3 Skipped Deployments

Project	Deployment	Updated
mecheval	Ignored	Oct 17, 2026 12:23pm UTC
vcad	Ignored	Oct 17, 2026 12:23pm UTC
vcad-docs	Ignored	Oct 17, 2026 12:23pm UTC

…the board brings its own fanout, one-dielectric via hops, raster-prefiltered via sites, cached pin fields The brain board's first pass: 619 connections in 195 s (it did not finish in 2 hours before). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

The brain board (619 connections, 0.4 mm BGA, 8-layer HDI): 2:09 at 73% routed, against >30 min (timed out) for the full arsenal at 74%. - --fast: no rip-up, negotiation, joint repair, CLI second pass, direct links or plane-stitch repair; legalization keeps only via merge and dead-copper pruning; a failed speculative search is not re-run sequentially. - drc::net_islands: island counts for many nets from one connectivity build; the CLI used to rebuild the whole board's connectivity per net. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>