Hull nesting: validation status

Nesting is enabled by default for hull macros, by the owner’s explicit decision on 2026-10-07. PNR_HULL_NEST=0, run.py --hull-nest 0, or a ladder campaign’s hull_nest = false opts out. Plain placements without hull macros are unaffected. The legality checks, native DRC gates, un-nested knit and rectangle fallback are unchanged. This default choice does not imply that nesting improves every routing metric.

Matched native example before the default change

The integrated head e5332b11625a7158ceb3333bc6ca521452370163 with --hull-nest 1 was compared with main c57077df3eb59bef496230558f330add15f3d7b0 using case 13-dovetail-blocks-23, seed 0 chosen before either run, the standard full search budget and automatic seed selection. Both used identical generated inputs and the same native maze runtime and headless KiCad 10.0.6. Both selected top-03-route.

Final metric

Main default

Nesting enabled

Hull interlock (mm²)

21.5

35.3125

Hull collision (mm²)

0

0

Macro rectangle overlap (mm²)

82.25

126

Vias

30

37

Copper length (mm)

524.764992

526.062197

Both passed native DRC with zero violations and unconnected items, preserved the netlist, completed knitting and satisfied all three constraints. Metrics were recomputed from recorded final macro poses and hull geometry. Main already interlocks; the nesting arm uses seven more vias and slightly more copper. This supplies an actual native-clean interlock example, not a global quality claim or validation of the later default-on head. The final-head versus main campaigns remain pending.

Two recorded native final boards

This comparison contains exactly the two recorded native final boards with a fixed viewport and no interpolation. Full logs, boards, traces, DRC results and provenance are preserved separately with the proof.

Campaign receipts

The following existing GCP summaries were fetched and checked on 2026-10-07. They ran source f990b165, before the integration with #88, #90 and #93. These historical receipts do not validate the default-on head.

Campaign

Coverage

Verified result

20261007-ladder-ff64b7

Eight ladder cases, seeds 0–1, base/nest

16/16 pass in each arm; matching verdicts, opens, vias and copper hashes

20261007-ladder-49f247

35 flat hard cases, seeds 0–1, base/nest

61/70 pass in each arm; the same nine failures, matching verdicts, opens, vias and copper hashes

20261007-ladder-b5f559

Twin-bank, seeds 0–3, base/nest/hp/hpnest

4/4 pass in every arm; zero final hull collision area

20261007-ladder-b4292b

Dovetail, seeds 0–3, four arms

Still queued, zero run time; missing C4D Spot instance template in northamerica-northeast1

20261007-ladder-9c11aa

Quad-bank, seeds 0–3, four arms

Still queued, zero run time; the same missing instance template

Twin-bank totals over four seeds:

Arm

Bounding box area (mm²)

Copper (mm)

Vias

base

5214.236

2130.173

164

nest

4083.859

2083.441

162

hp

5006.464

2264.873

152

hpnest

4409.349

2208.681

155

The missing dovetail and quad-bank comparisons remain an acceptance gap. The completed cells do not establish that every hierarchical rung is no worse. Replacement jobs require the final tested source and the approved total spending cap.

Animation evidence

Accepted poses of the synthetic half-macro fixture

The committed animation runs the existing half-macro placement fixture through hull.nest and draws only its recorded accepted poses, with a fixed viewport. It performs no interpolation, routing or native DRC. It demonstrates legal macro-box overlap; this simple half-macro fixture is not evidence of true L-shaped hull interlock. Reproduce it with the engine dependencies and Pillow 10 or newer:

PYTHONPATH=hardware/pnr:hardware/pnr/tests python tools/animate_hull_nest_fixture.py \
  --out docs/feature-animations/hull-nest-fixture.gif

The original local 13-dovetail-blocks-23, seed 1 storyboard was also inspected. It comes from the recorded run at 03c26c013e3303ce21e1342450deafe098ffe568 with --hull-nest 1 --maze-kernel packed. It is native board evidence, not a completed GCP A/B.

The recorded final result passed: zero opens, zero native DRC violations, 31 vias, 421.012 mm copper. Final macro rectangle overlap is 8.5625 mm², but hull interlock is 0 mm² (and hull collision is 0 mm²). Therefore this animation does not demonstrate the intended true dovetailing acceptance criterion. It is deliberately not presented as evidence that the missing campaign passed.

Integration checks

The focused Bazel targets hull_nest_test, macro_hull_test, gp_polish_test, hard_rungs_test and tests/unit/exp:test_kinds passed against the integrated tree. Full CI still needs to pass on the pushed head.