Worklog¶
A short, live status board: rewritten at the end of each session, not appended to. History lives in git and in the pull requests.
Last updated: 2026-10-11 (USB paired-routing integration validation).
PR #99 merged into main at 5d1a07c1; its engineering workbench and manufacturing
follow-ups are available. PR #94 is rebased on that revision.
In progress¶
PR #94 integration: corrected compact recovery’s offset-courtyard geometry. Fresh Linux A/B passes all 12 USB cells under control, prior-paired and integrated defaults; both local integrated header reruns pass native gates. Total uncoupled length remains 2 mm. Validation receipt preserves all baseline/repaired attempts, root/candidate seeds, selected starts, settings, pins and saved-board hashes. Accepted native header snapshots illustrate recovered placement. Full repository checks and hosted validation gate landing; wider board/topology coverage remains open. Bounded search acceptance does not require every root seed to pass.
Assembly inventory: quantity-aware catalog filtering and assembly-house checks publish replayable BOM/identity/quantity evidence. JLCPCB public inventory includes orderable stock and declared allowance/minimum screening; PCBWay explicitly needs its own supplier confirmation. Native Manufacturing exposes per-part shortages and historical checks. Network failure/unknown stock never passes sourcing.
Manufacturing GUI: approved PCBWay packets have an explicit human-clicked official-plugin file upload and quote handoff, with exact artifact receipts. JLCPCB automatic attachment awaits approved API access; manual handoff remains.
Manufacturing package discovery: published native/prototype packages are discovered automatically. Supplier, quantity, finish and optional through-hole service lead to local vendor-specific BOM/CPL preparation, previews, content-bound review and human downloads/quote handoff. Prototype vendor DFM/rotation and physical verification remain explicitly open.
On-demand parts: authoring CLI builds fetch missing component queries, retain response digests and selected assets/locks in the article, and replay offline/frozen. Operator our rules_atopile-derived picker matches public supplier facts without hosted Atopile calls, accounts or tokens. No shared catalog uploads or prepopulated local library are required. Typed discovery covers resistors/capacitors and explicit LCSC/MPN picks; network/data failures remain explicit.
Workspace search and Performance: Cmd/Ctrl-F opens a floating source/component/net search, including in embedded frames; repeated shortcuts preserve/clear queries. Inspect tab layouts migrate to the dialog. Performance holds native controls and validated project cloud settings with read-only SDK status and no job submission.
Recorded build attempts: Experiments lists archived schematic-build results, including failures before PnR, and opens their result JSON in read-only source tabs. Restored active tabs initialize on first drawer open. Embedded styling and saved theme apply before renderer scripts load, avoiding the legacy-theme flash. Live discovery is read-only and retains failed attempts. The workspace server accepts cold-start asset bursts without dropping concurrent proxy connections.
Experiments embed: workbench tabs own panel visibility; the native legacy drawer controller is bypassed in embeds. Experiment lane selections follow through to the live board/schematic/3D views while pinned comparisons remain independent.
Workbench follow-up: drawer open/dismiss motion respects reduced motion; startup no longer waits for provider discovery. Source syntax, import navigation and hash-scoped selected-text Ask context pass browser checks. A project directory tree starts in the left drawer.
Prompt-to-PCB agent workflow: owner-provided engineering workflow is now scoped into AGENTS.md; the installed wheel ships exact instructions and execution guide.
agent initpreserves contracts/checkpoints,agent instructionsexposes guidance andagent chatopens an operator-owned interactive Codex/Claude terminal. The image includes checksum-pinned Codex and Git; no credentials or paid model sessions run in CI. Focused bootstrap/installed-wheel/container checks and hosted validation are in progress. Complete populated RF assembly acceptance remains #97, not satisfied by DUT-only packaging smoke. Review-first requirements/risk presentation, bounded speculative implementation, refinement-driven restarts and inline visual feedback are now explicit in both embedded guidance and the launcher prompt.workflow init/query/nextimplements a declarative, evidence-gated iteration controller with atomic state/history, immutable receipts, revision invalidation and acceptance/completion guards. Receipts remain assertions requiring genuine user decisions and authoritative validation reports; the controller does not authenticate them or sandbox agents.Workspace-native browser UI: replaced the OpenCode overlay with one layout for projects, focused/main conversations, artifacts, engineering views and a resizable chat pane. Background activity collapses into one spinner/details group; the footer follows controller stages, including review and rework. The journaled main-thread harness continues idle agents with their selected model, stopping at genuine review checkpoints and native questions/permissions. Turn duration, elapsed session time and continuation counts never stop the design loop. Stop and explicit resume are persistent. Runtime logs remain exportable but are excluded from Git snapshots to prevent recursive chat-summary growth. YAML requirements/risk/trace models now populate a native browser using pinned rules_requirements semantics. Report artifacts link to owned verification cases; stale or corrupt evidence remains open. Content-bound artifact review carries canonical feedback/questions, explicit approval and a canned revision cycle. Bundled package inspection is allowed without repeated external-directory prompts. Schematic capture opens live atopile sources alongside a netlist-backed schematic, retaining the last valid export through incomplete writes and failed builds. OpenCode remains the headless provider/tool/session runtime. Canonical notes carry immutable focused transcripts into the main conversation without starting a model turn; shared document edits use revision checks. Pinned sandboxed three.js choices, rendered artifact markup, full conversation journaling and integrity-checked workspace archives are implemented. Browser checks distinguish synthetic review fixtures from actual engineering views. Solver regeneration remains unverified until a pinned replay recipe passes; image-build toolchain warming remains outstanding.
External RF packaging (#96): collect every RF Bazel subpackage explicitly without numerical dependency closure in the wheel collector. Export/report, planar, Palace, coupons and order0 modules now ship, with the authoritative Palace schema and registered coupon copper/results. An isolated installed-wheel test runs design/export/native revalidation outside the checkout; both image architectures run the same external-project smoke. The installed-wheel smoke passes; full-suite and two-architecture CI validation are pending.
PyPI publishing: the release workflow publishes the same two tested Bazel wheels through OIDC, with version/project/architecture checks and no dry-run uploads. First upload requires the owner to register the documented pending trusted publisher.
RF CI lanes and Bazel cache snapshots (branch
codex/rf-nightly-ci-cache): expensive RF solver/gradient/design and coupon fitting coverage moves to nightly/manual Linux and macOS; ordinary CI keeps inexpensive RF units and a deterministic native optimizer smoke. Expanded Bazel targets are checked against an exhaustive lane inventory. Explicit cache saves retain completed outputs after test failures with fork PRs restore-only. The existing CI dispatch can explicitly invoke the RF lane before merge. Native smoke, 14 focused policy regressions, changed-file hooks and actionlint pass; both hosted platforms verify the expanded 65-target inventory. Both RF lanes and same-head cache replays pass; detailed measurements and combined-head CI status after #91/#92 are tracked in PR #95. Lane inventory and policy.Hull nesting default (2026-10-07, PR #91). Owner requested default-on nesting for hull macros with explicit opt-out. Legality/native gates and both routing fallbacks are unchanged. The integrated native dovetail example is clean with nonzero interlock; vias increase in that one pair. Final-head versus main dovetail and quad-bank campaigns remain pending under the approved total budget. See validation status.
USB matching recovery: default-on runner reorderable line slots with joint common half-turns, off-center/split meander windows, bounded transactional grid-route reconciliation, and a saved-copper free-space adapter gated by cold native DRC. Final two-case/two-seed A/B: baseline 2/4, free-space windows 3/4, corrected line search 4/4, line search plus sequential recovery 4/4. Every passing finalist has zero native DRC violations/opens and legal placement. Exact original MCU board also passes without moving foreign copper. The reviewed resistor-facing gap was reproduced and fixed: a joint half-turn/order reversal keeps both board slots and saves 7.30 mm pad HPWL; cloud saved resistors are both at 0 degrees. All 288 Bazel tests and all pre-commit hooks pass. Broader ladder validation remains necessary for wider coverage. General coupled-route pushing remains unfinished.
Joint differential-pair access: required coupled routing enabled by default in the regression runner, with
--no-pair-accessfor the control, joint through-via site/orientation selection, bounded shared-surface/alternate-layer search, whole-run length/delay compensation and a saved-copper coupling gate. Two USB cases at seeds 0 and 1: control 4/4 and coupled arm 4/4 native-clean; all eight declared pair segments in the coupled arm pass coupling, with at most 1.522 mm uncoupled copper per leg against the unchanged 2 mm limit. Independent legs count as unresolved for placement selection. Branched pairs and mid-trunk transitions remain unsupported; wider ladder validation remains to be measured. All 298 rebased Bazel tests and pre-commit hooks pass; new audit/tuning stages map into the existing live timing vocabulary. Rebased bare-default A/B (line/window options off): control 2/4, strict default 0/4; three default boards are native-clean but fail saved-copper coupling, and header seed 0 also fails an open/skew gate. The default remains enabled by explicit designer request. The designer subsequently requested enabling line reordering and window search alongside paired access, matching the 4/4 arm; all three runner options now default on with explicit opt-outs. Sequential recovery remains opt-in. PR #94 includes both placement and routing changes. Linux CI’s whole-tree privacy test hit its 60s small-test timeout without a finding; classify that test as medium (bounded 300s), preserving every check.Route-then-compact and hull packing (PR #86, branch
claude/route-compact; design docs/design/route-compact.md).PNR_ROUTE_COMPACT(TOP,BLOCK,FLAT; default off): after routing, order-preserving 1D compaction (x then y) by the copper each gutter holds, rip-up and reroute, back-off on a worse route; the hierarchical driver measures block hulls from its in-memory routes (PNR_MACRO_HULL), GP packs hull bodies (PNR_HULL_DOVETAIL), new rung13-dovetail-blocks-23. Rebased on main (#80’s viewer stage timing wraps the reroute step now). Review-fix round (2026-10-06):not_worsecounts deferred nets andport_debt_mm, not just missing/unresolved; its vias/ copper tolerance is a share of the pass’s original route, not the previous accepted step (no more ratcheting);feedback.pyrefreshesdeferred_nets/convergedafter FLAT instead of leaving them stamped from the pre-compaction route; the pass skips outright when the base route is itself incomplete; the deadline’s first-reroute estimate no longer borrows from wall time spent before route-compact started;compact_blockcounts violations, not kinds, and gains the flat board’s legacy-plane/coupled-pair guards; softedge_align/regionrefs hold during compaction; newpnr.place.hull.macro_overlap_area_mm2metric andhier["hull_rects"]for renders to draw real hull geometry. Hierarchical A/B rerun at head in progress (campaign20261007-ladder-01bca2, adds a hulls-only-no-compaction arm). Next: rebase on #64/#70 and A/B again on top of detailed placement; hull support incompact.pyand the keep legalizer (gp_polish still turns off over hull bodies); the push-and-shove router as the reroute.Energy track optimization (#90, default off). The witness-growth/reverse-cost proposer is wired into the native gloss transaction controller with exact native checks, joint-refill IR and reference evidence gates, protected metadata/copper guards, spatial quiet-chain invalidation, and phase-end rollback. Shapely/GEOS dependencies are hash-locked separately for controller and KiCad Python ABIs. Focused tests and actual clean-fixture native acceptance/refusal/byte-identity checks pass; a four-state recorded native animation documents the change. The KiCad dependency install now uses root’s build cache and copy links, preserving runtime-home ownership; a focused regression checks its environment. Image build/smoke and Linux aggregate CI pass. The macOS ABI-path test now checks only the canonical path is added; all 16 integration tests and eight hard-edge tests pass locally, the latter within its unchanged 60-second limit. Next: fresh macOS CI, then review before merge. RF/SI-dependent plane edits without complete external evidence remain refused; no radar-board completion or joint pair solver claim.
Routing engine for radar60 Rev A, stage 3b (#49, branch
claude/radar-routing-engine-2; all opt-in, byte-identical where undeclared). Router: pad-local clearance and mask margin (data),board.class_clearance: maze | repair,board.dru_routing(a board’s.kicad_druwhere it constrains routing; the rest listed as unmodelled),board.edge: exact, and SMD lands on the far side of their footprint (Pad.far_side). Power: partial fanouts (the fanout’spartial: bridge, retry),plane_partition(connected per-rail territories on a power layer; trees only where a zone of the minimum width fills, foreign copper at the pair’s class clearance) andir_drop(a resistive network per rail on the refilled board;unsolvedwhen CG stops short; the solve runs inPNR_PYTHONwhere KiCad’s Python has no numpy). New hard rungs11-ufbga201-...-classes,-partial(run.py excuses a spec’sdesigned_openpads, held exact by itsunconnectedcheck) and-rails. Identity regression againstmain(GCP C4D,--compact --gloss, 78 shared cells): placed, routes and rules identical in all; boards identical modulo segment direction in 77, the 78th (-blocks1) differs after the gloss stage only (its pre-gloss board is identical; KiCad’s dangling check on a 0.035 mm segment depends on the direction the gloss writes); 78/78 same verdict. Next: the radar trial’s open items (the U2 power stage escapes, LVDS as coupled pairs (A8), 1V2/1V8 delivery, the PA corner with macro v2).RF microstrip inverse design (#29, branch
claude/rf-topopt; design docs/design/rf-topology-optimization.md, guide docs/rf-inverse-design.md, choices in docs/decisions.md). Done: the solver (yapnr.rf: Yee FDTD, CPML, line ports, exact adjoint), the optimizer (material grid, conic filter, tanh projection, Zhou length scale, own MMA with the epigraph, robust variants, adaptive moves, specs with lumped resistors and their absorbed share, checkpoints, the best binarized design exported), net-tie.kicad_modexport, Touchstone, result JSON, the animation, and five end-to-end cases re-validated from the footprint on three grids. Round 2 (2026-10-02/03): a copper-edge correction and a modal port source, the antenna grown from the feed line alone, the cases re-run. Review fixes (2026-10-03, design §24): the footprint follows the pixel boundaries (the chamfered export had put other copper on the finer grids: most of round 2’s 0.7–2 % grid shifts), the repair widens the one-pixel necks the exact copper shows, the time-step library holds the diagonal patterns, thetrust_referenceoption, one thread per Bazel RF test with long timeouts (CI had timed out; it passes now),seed_overlapand the −10 dB band invalidation.json, the antenna judged over 9.7–10.3 GHz at η ≥ 0.7. Results (same copper on every grid): divider, combiner (topology given by seed and keepouts) and diplexer (a closed-form stub filter refined, 77 % seed) pass on all three grids; the antenna meets |S11| and η on all three but fails the 4 % power balance at 9.7 GHz on the finer two (4.6, 4.7 %; cause not established, design §24.7); the bank fails (adjacent channels −9.8 and −11.4 dB against −12 dB). A diplexer from the plain junction (robust, 25 iterations) only rolled off (rejection 13–16 dB; stopped). Artifacts indocs/rf/. Native kernel (2026-10-03,claude/rf-kernelsmerged withmain, docs/rf-solver-backends.md): the C stepper is the default backend wherever its library loads (auto, float64, bit-identical to numpy; numpy otherwise, said once), with every round-2 option (sha256 matrix intest_native_identity); Bazel builds it into//yapnr/rf, the wheel is per platform (manylinux_2_34 x86_64 and aarch64, macOS arm64) and the image loads it from the wheel (smoke-tested). Verified again independently (2026-10-04): the full-grid identity matrix on a C4D, end-to-end runs on the smoke grids, the divider and antenna re-validated on native float64 (round 2’s verdicts, within 2.5e-4 dB); one iteration on a C4D-16 takes 1.47 s (divider) and 16.1 s (D1 size) at 16 threads, 8.4 and 6.3 times faster than torch float32. Fixed then: a case’s run directory resumes with its own spec, a process solves each port mode profile once (the solve’s last bits varied between processes on the Mac), and the library’s first load is thread-safe (threads asking while it ran had fallen back to numpy). Next: the owner’s decisions (balance criterion, the seeded filter banks, the native defaults), sub-pixel tuning of binary copper for resonant filters, a modal port extraction for radiators, ayapnr rfCLI, an external cross-check, footprints in PnR.Compact placement (
PNR_COMPACT, shrink-to-fitPNR_SHRINK, both off by default) (branchclaude/compactonmain; design docs/design/compact-placement.md): offset courtyards, the compact legalizer, spread 1.0 with clustered starts, a compactness tie-break after the vias, legacy plane drops planned before routing (DROPS) and the flat driver’s outline search. Flag-off identity holds on x86 (placed and routed bytes of 30 cells againstmain). The A/B (GCP, seeds 0 and 1) passes the ladder, the showcases, the header rung,08-chaser-20-planeon ten seeds and the nightly hard rungs, but 5 of 16 manual09-mcu-usb-31cells regress (USB pair skew), so it stays opt-in. The docs and README animations (ladder and showcases) are regenerated with--compact --gloss(renderer 3: the gloss stage as a before/after; CI’s traced runs take the same options). Next: reserve room for pair tuning under compact, then another A/B.Hard-rung gap fixes (branch
claude/gap-fixesonclaude/ladder-hard-rungs, local, not pushed). The hard rungs (regression/hard_rungs.py: stackups, via kinds, sides, absolute and relative constraints, Monte-Carlo search, a THT header) found capabilities the engine lacked. Five tracks are merged, each new behaviour following from the board’s inputs (a board without them routes as before; the eight ladder cases and four showcases give identical boards):declared stackups (
pnr/stack.py): a KiCad stackup with apowerormixedlayer, or a custom rule keeping tracks off a layer, routes every signal layer of any count; plane drops are planned with the pin escapes; every plane layer is formed (full outline by default);blind, buried and micro vias (
pnr/via_policy.py, the rules’via_policy): only spans the declared stack can build, one build per board, return ties between two planes of one net;sides (
board.sides: double,pnr/place/sides.py,detail_moves.py): placement chooses the side of free parts; writeback mirrors bottom parts onB.*;length matching (
pnr/route/detail/tune.py,pnr/length_model.py,pnr/place/matched.py): declared pairs and groups are tuned with meanders against KiCad’s own length measure;router speed: the native kernel (the packed search over dense per-net fields, in C; identical routes; the packed Python kernel where no library loads,
PNR_MAZE_KERNEL=packedon request) is the default (2026-10-07: a 192-cell GCP A/B, every route identical, -54% task time), and an exact-separation recovery (PNR_EXACT_SEPARATION=recover, the default) for routes the halo model leaves open.
Results (two seeds each): ladder and showcases 24 of 24, CPU 2,281 s to 442 s; hard rungs clean in 37 of 48 runs (7 before), CPU 11,537 s to 1,964 s on the 46 runs with a before. Still failing: 07/09
absandrel(region,align: onclaude/gap-constraints, not merged), 09header(no legal placement) and 09 4L-SSGS seed 0 (USB pair skew after a detour). Three “Ladder fix” commits came with the tracks (the plane check, the side check, the HDI microvia size).Gloss, dekink and corridor coalescing (
PNR_GLOSS, off by default) (branchclaude/gloss-port; design docs/design/gloss.md): ported from Splanc’s src18 with its tests,board.Deleteinstead ofRemove, workers throughpnr.proc.run_status; the06g-glossand07g-glossloop passes, the ladder’s opt-in--glossstage with--gloss-flagand--gloss-measure, functional groups derived from the rules (PNR_GLOSS_CLASSES_FROM, the most specific group wins) and the public example groups file for07-chaser-20. Review fixes: the ray metric no longer depends on which way a segment is drawn (patch handed back to the source), gloss settings in the router key and imports, the code key covers the pass only with the flag, the ladder’s outer gate by violation keys, frozen and hashed groups file, per-stage CPU as on the hard-rungs branch, a nightly CI lane. Identity with the source (plus the same ray patch) is exact on three private replay points; flag-off identity withmain(stubbed loop) holds; a real native loop with the flag ran both passes on a public board. Ladder A/B (24 runs per arm, no regression, bends -32 %, length -4 %, CPU 2.2 times): see docs/regression-ladder.md. Next: the larger paired native-loop A/B on the cloud lane decides default-on (owner); rebaserun.pyontoclaude/ladder-hard-rungswhen it lands.Fab outputs and staged ordering (branch
claude/fab-order; design docs/design/fab-and-ordering.md, guide docs/fab-and-ordering.md). Built: vendor profiles and stackups as data (oshpark-2l/4l/6l,jlc-4l; draftsjlc-6l,pcbway-std,pcbway-hf-2l), engine support for them (PNR_FAB_PROFILE,run.py --fab-profile;legacyandjlc-pofvunchanged, golden-pinned),yapnr fab profiles|show|check|build|previewandyapnr order stage|vendors(staging only, never uploads). Ladder cases 04 (oshpark-2l) and 08 (oshpark-4l,jlc-4l,pcbway-std) route clean under them and stage as dry runs. Not built: O2 and O3 (decisions D1 and D3), JLC rotation corrections in the part cache.Merged today: PR3a (#9, code key scheme 2 and the engine format), the viewer (#12, PR4:
yapnr/viewer,bazel run //:viewer -- --root <live>, paid features off by default, elkjs and three.js fetched pinned; guide docs/viewer.md, choices in docs/decisions.md), and #13 (an explicitstarted.jsonstamp for the routing-feedback staleness check, fixing #10 on Linux). The format commits of #9 and #12 are in.git-blame-ignore-revs(#11, #14).Merged since: the regression ladder in CI and the PnR animations (#15:
pnr.trace,pnr.provenance,pnr.animate,ladder.yaml,docs/regression-ladder.md) and two viewer test race fixes (#16, #17).atopile toolchain and part cache (merged as #18; A1 and A2 of the end-to-end plan):
yapnr atopile setup(hashed per-platform locks, atopile 0.15.8 on Python 3.14.7),yapnr atopile build(offline, isolated, bounded; a hook in every atopile interpreter), the loopback picker (catalog schema v1), the part cache (store, HTTP server with tokens and takedowns, clients, importer, container), the Bazel toolchain andyapnr_atopile_build. The adversarial review’s findings are fixed: a lock pins its part; client tokens go only to the named server (no redirects, https or loopback, the write token never with reads); the server checks tokens before bodies, bounds its connections, serves only files a part uses, collects orphans, checks file types, and a takedown blocks the part’s files; local-only parts; no git clone offline and no contact with a running KiCad; ato.yaml’s layout paths; the Bazel action copies only its declared inputs and needs an explicit cache; hashed sdist builds on linux-aarch64. Real atopile: 4 end-to-end cases (7 builds) pass on darwin-arm64;bazel test //...137 of 137. Not done: the image’s/opt/atopile(A4) and a KiCad 9 against 10 A/B (docs/frontends/atopile.md).Splanc’s parts in a local part cache (local-machine note; Splanc is not modified): the committed parts of
splanc,splanc_dev,splanc_maxandsplanc_eol_tester(276 directories, 248 versions of 163 parts) and the 103 entries ofpicker_catalog.json. They are in the development Mac’s default cache,<part-cache>=$XDG_DATA_HOME/yapnr/part-cache. One lock per board is in$XDG_DATA_HOME/yapnr/splanc-locks/. Neither is in any repository. Each part recordssplanc@<commit>:<path>, its LCSC id, its generator (easyeda:C<id>or hand-authored) and a licence note.splanc_max’sOUTalso has a version with the git-ignored DEGSON STEP model (splanc_max.local-cadlock), markedlocal-only(never uploaded; a public server refuses the cache). Every lock materializes the committed files byte-identically (rechecked after the review fixes, also over HTTP). Mini, built from a scratch copy with its parts directory emptied, has the same input id as a build from its committed parts, both from the directory and over HTTP, before and after the review fixes. Seeded with the same local layout, it equals rules_atopile’s last Nix-built board in nets, footprints, pads, positions and outline.splanc_eol_testerfails on 0.15.8 with or without the cache: atopile rejects itselec/footprints/directory as deprecated.Constraint and hierarchy showcases (branch
claude/animations-groups-hier, on top ofclaude/ladder-animations; designdocs/design/constraint-and-hier-animations.md, §10 and §11 “As built”). Engine, all opt-in: a HARDline_groupconstraint placed as one rigid macro insideplace(),edge_alignhard/tolerance_mm, own-net fixed copper in the detail router, andregression/hier_case.py(blocks placed and routed on their own boards, placed as macros, knitted with block copper held fixed). Traces gaingroups,group_members,header.constraints, afixedevent and per-template block traces.run.py --showcasesand--trace-placement-every;designs.showcases():line-chaser-20,edge-io-12-free,edge-io-12,hier-twin-bank-32. Showcase run (pool seed 0, 8 starts, 3 finalists, snapshots every 5): all five cases pass the gate and the constraint audit, 0 opens, 0 findings; vias 19 (07), 20 (line), 9 (edge free), 8 (edge), 38 (hier). A/B of the eight ladder cases: identicalplaced.json,routes.jsonand trace digests. Renderer: constraint highlighting (pnr.animate.highlight), rigid-body tweens, side-by-side comparisons (python -m pnr.animate --compare A B,pnr.animate.compare), the hierarchical chapters (pnr.animate.hier,provenance.from_hier),--pacing showcase, and//hardware/pnr:showcase_animations. Four files indocs/animations/(13.1 MB; folder budget 30 MB, the hierarchical WebP 3.5 MB), the page docs/constraints-and-hierarchy.md, a second README item (the side-by-side chaser GIF), and a nightly, informational showcase step inladder.yaml(never run yet). Re-rendering the eight ladder cases reproduces every committed file’s SHA-256; each showcase renders byte-identically twice. Review fixes (design §12): line-group macros reserve their members’ sides (line:<name>) and refuse source-locked members; knit retries trace in their own scopes;PAD_AXISis checked against the footprints; half-turns are drawn as flips, off-board poses widen the camera, the edge comparison replays its pool’s eight starts; GIF palette and per-file renders. Rerun.yapnr/ladder/showcase-2(enginec4db5fb): the five cases’placed.json,routes.jsonand trace digests are identical to the first run; all pass. Next: the owner’s review of the decisions indocs/decisions.mdand the pull request.RF fab-model coupons (#32, branch
claude/rf-coupons; designdocs/design/rf-fab-coupons.md, guidedocs/rf-fab-coupons.md):yapnr/rf/couponsgenerates boards A (JLC04161H-7628) and B (JLC06161H-7628), DRC-clean, with expected S-parameters and fab zips underexamples/rf-coupons/, and extracts the fab parameters from a VNA session (python -m yapnr.rf.coupons extract): multiline TRL, a joint fit with connection nuisances and a parametric bootstrap. Synthetic recovery passes in CI. Next: the owner’s decisions of design §14 (connector, VNA, scope), checking the SMA footprint and JLC’s edge clearance, and the order (owner only).RF coupons, adversarial review (#32): the boards’ TRL sticks carried no stick id (fixed, all labels now
A05 P L40); mouse-bite tabs widened to JLC’s 5 mm minimum; multiline TRL dropped nearly degenerate line pairs, whose swapped roots spiked the error boxes and failed a quarter of nominal verification lines; a fit stuck at the 2D-table edge is now a reported problem; board B alone does not split the core’s and the 7628’s εr (use board A’s fit as--prior). Impedances re-checked with an independent finite-volume solver (within 0.1 Ω) and IPC-2141A (stripline 50.5 Ω); recovery re-run with another seed, a wrong nominal stackup and a worse lab. Open: the SMA part, the panel fee, a mask-off stick on board B.RF coupons, Order 0 on OSH Park (branch
claude/order0, fromclaude/rf-coupons; pages docs/rf/order0/): the Cinch 142-0701-851 launch of regions M and W (2D-designed), board O as three uploads (O0-M coupons + R1 + A16 + A04R + tag/QR; O0-W light with the D2 window; O0-D with the D1 window), the OSH Park frameless panel, line tables for FR408HR (EM528 shares them), coupon predictions on both substrates; board B moved to JLC06161H-2116C with a mask-off tie line. All DRC-clean;yapnr fab check0 errors. Next: D1/D2 copper and the R1/R1t FDTD predictions (#29 round 2), the fab bundles, the pre-registration release, the capture tooling (WP7).Order 0 part 2 (branch
claude/order0-p2, main +claude/order0): RF stage plans require the native FDTD library and passYAPNR_RF_THREADS, backend, dtype andYAPNR_RF_*per job,--image-commitrefuses a bundle whose C sources are not the image’s (PR #30 review follow-up);yapnr.rf.coupons.equivalent(2D thickness-equivalent substrates of regions M and W on FR408HR and EM528, matched to the coupon model’s line at 5 GHz);yapnr.rf.order0(D1/D2 specs in four formulations, the selection rule, criteria, forward-run directories of R1/R1t and of run 0b’s loss lines, the loss correction and D-O0-11); docs/rf/order0/demos.md and the generated inputs. Run 0b and the D1/D2 compute stage on GCP Batch (C4 Montreal, C4D us-west4);d1-starpasses all three validation grids and is merged into O0-D’s window (catalog.o_optimized, one real DRC clearance call left for the owner); no D2 formulation passes (closest miss 0.012 dB). openEMS-vs-yapnr.rf predictions collected (D-O0-13a); pre-registration bundle and manifest indocs/rf/order0/. Open: the D1 island/DRC and D2 near-miss decisions, the release/OpenTimestamps step (owner).Palace, the second RF solver (#46, branch
claude/palace; guide docs/rf-palace.md, imagedocker/palace): planar documents (yapnr.rf.planar: lines, the radar60 patch and feed models, KiCad regions with footprint pads; lossymetalfloor; solder mask over solid copper), the Gmsh builder, configurations checked against Palace’s schema, the validated sign-off settings (python -m yapnr.rf.palace case),yapnr.rf.palace.results, andpalace_plan.py/palace_job.pycampaigns with stages. Validated against openEMS and the 2D cross-section solver on the radar60 lines, single patch and TX1 feed; two reviews’ fixes in. Open: Palace issues to report upstream, a rank-scaling run before the bank model, the image’s ParMETIS licence (owner).
Integration checks¶
PR integration review (2026-10-07). PR #92 is rebased onto main after #88 and #90, retaining only its two new optimization commits. The existing 96-cell A/B receipts were rechecked: 96 matching verdicts and 94 matching copper hashes, 85 passes per arm. See validation evidence.
Next¶
Owner (viewer, #12): review after the fact; decide the agent’s default model (opus, $2 per turn, $20 per process) and file the issue for
pnr.capacitor_intent(the cost replay reports contexts that need it as unavailable).PR4 follow-ups:
--projectand the manifest’s[viewer]table (with PR3d), the publishedyapnr-live-event-v1JSON Schema, browser e2e tests (Chrome DevTools), the viewer in the wheel and images once the engine is (PR3b),examples/led555as the smoke run (PR6b).Before importing trials made by a yapnr checkout from before PR3a’s format (legacy code keys, no
code_key_scheme), keep a frozen copy of that pre-format tree (they re-key from it) or import with--import-code-mismatch rebaseorwarn; the defaulterrorrefuses them. Splanc’s frozen snapshot trees are not affected.Owner: A/B Electrical221’s two flags on the hierarchical engine before turning them on. Before the next experiment runs this engine, one rung-1 evaluation with it (
board.Deleteeverywhere) when the Mac is free: the replay covered 5 of the 27 changed sites, and a full evaluation’s nested workers exceed the two-KiCad-process budget kept while H7 runs.Owner: push Splanc’s
splanc-mini, so the 9 newestImported-Fromlinks of PR1 resolve (see the manifest).atopile follow-up PRs (docs/frontends/atopile.md, “Not done yet”): A4
/opt/atopilein theyapnrimage (the aarch64 wheel builder stage, notices, size), a KiCad 9 against 10 A/B and an end-to-end project that uses a stockLibrary:Footprint, then A5 to A8; after them the spec and checks (S1 to S4), the agents (R1 to R3, off by default) and fab outputs and ordering (F1, F2, O1 to O3; ordering stays staging-only: cart, quote or payment page).Owner: move Splanc’s builds to the part cache. Commit each board’s
yapnr-parts.lock.json(from the local locks), build withyapnr atopile buildoryapnr_atopile_build, and only then remove the committedelec/src/parts/. Also commit or decide onelec/layout/(it seeds designators), and movesplanc_eol_tester’s footprint into a part.PR6a, then PR3b onwards: wire the 67 unwired engine test files with the glob macro (the two hygiene tests and PR3a’s
test_feedback_signalsare wired; the two feedback generation tests skip without Splanc’s Mini inputs) and replace the Splanc defaults and fixtures (manifest, “Known leftovers for PR3”). The KiCad-dependent tests run under the headless KiCad Python only (DEVELOPERS.md). PR3a landed before the KiCad lane: its format is checked by syntax-tree equivalence and fresh-interpreter imports, not by a KiCad-side test run. PR3b burns down the.flake8baseline and fixes thevia_coalesceF821 with a test of its own.PR-R is merged (#2). Its first
mainbuild pushesyapnr-kicad:10.0.6-1-src, thenyapnr-kicad:10.0.6-1(with10.0.6and10.0), thenyapnr:edge. Then make both GHCR packages public (package settings > Change visibility; irreversible, owner-approved). The organization must allow public packages first (Organization settings > Packages > Package creation: Public), or the option is missing. Check that an anonymousdocker pull ghcr.io/studio-fug/yapnr:edge(andyapnr-kicad:10.0.6-1-srcon arm64) works and thatgh attestation verify oci://ghcr.io/studio-fug/yapnr:edge -R Studio-Fug/yapnrpasses.Create the release-note labels (
tools/release/create_labels.sh), label open pull requests, and run the release dry run once (gh workflow run release.yaml --ref main); read the notes preview in its summary.Uncomment the container and release badges in
README.mdonce the image is public and v0.1.0 exists; decide on immutable releases (docs/releases.md).PR6a: the KiCad-side workers run under
/usr/bin/python3(3.12), which cannot import yapnr (installed in the 3.11 venv only). Give them an import path with yapnr’s pure-Python modules and none of the venv’s compiled packages, and add a smoke check for it.First release tag
v0.1.0(owner) once the engine runs end to end inside the published image on both architectures (the PR6b example), after theImagerun of that commit onmainis green; make theimagecheck required then.Owner, fab and ordering (design §12): decisions D1 to D5. For the first OSH Park order, route the board under an OSH Park profile, run
yapnr order stage BOARD --vendor oshpark, drop the zip on oshpark.com, check the preview and pay there; RF boards also pin--stackup(D4).Owner decision (2026-10-07,
docs/decisions.md): thejlc-pofvdefault flip. All 21 of20261006-ladder-b0ec34’s jlc-pofv failures pass onclaude/rails-settle(the-railscell was a plane-partition bug, fixed there), and the 45 gated cells pass 90 of 90 underjlc-pofv(88 underlegacy), but the manual target rungs leave more open under it (11-shove-channel-lm-143 vs 2,12-soc-bga-11351 vs 33 and 35 and one 3600 s timeout). Stayslegacyuntil the owner says whether those count for “no rung worse”.Rebuild the KiCad base monthly (bump
docker/yapnr-kicad/TAGto the next-N), or with the Dependabotubuntudigest update (docs/releases.md, “Maintaining the images”).Showcases (
claude/animations-groups-hier): the first nightly showcase step on the arm64 runner (its runtime and its placements are unmeasured;showcase-edge-io.webpfits its 2.5 MB budget at quality 70 by 450 bytes here, the encoder steps down where it must). The ladder’s own animations still sweep single parts through half-turns (their timelines are pinned); move them to flips with the next deliberate refresh ofdocs/animations/.Hard-rung gaps: merge
claude/gap-constraints(region and align; it conflicts with the sides track in seven placement files) and rerun theabsandrelrungs; give the placement model an origin-to-courtyard offset (the THT header’s origin is pin 1, so its courtyard is placed off by half its length and never legalizes); carry the three Ladder fixes (plane check, side check, HDI microvia) toclaude/ladder-hard-rungs, or land the rungs and the fixes together.Done (2026-10-07, owner: flip to native if no rung is worse and routes are equivalent): the C maze kernel is the default where its library loads (Bazel, the wheel and so the image, or a host compiler); CI runs
dense_maze_native_testandexact_route_native_testin//....Boards with blind, buried or micro vias: the packed and native kernels hand them to the reference kernel (so no exact-separation recovery there), the length tuner only adds meanders, and the native KiCad repair loop and the hierarchical driver add through vias only.
Blockers¶
The Docker VM on the development Mac has a full disk (158 GB, 100 %), so the linux-aarch64 container run of the atopile review fixes (wheel build,
setupwith the hashed sdist-build lock, end-to-end tests) and the part cache image smoke test could not run; the images there belong to other projects and were not pruned. The no-isolation sdist build was checked with the pinned CPython on macOS instead.
Known issues¶
detail_route_test(large, not in the PR lane) failstest_drc_clean_by_constructionon the Mac at the branch base as well (a shared footprint cell on thesplanc_devfixture); not caused by this branch.
Do not retry¶
Exporting RF designs as marching-squares contours of the pixel centres (half-pixel chamfers): on finer grids that is other copper than the optimizer’s, and the re-validation’s resonances moved 1–2 % for it. Trace the pixel boundaries (
yapnr.rf.export.contour).Widening a one-pixel neck at whichever facing void pixel comes first: next to other copper it leaves a one-pixel gap that the space pass closes, merging islands (the antenna’s matched band fell from 19 to 4 %).
export.repairchecks the distance to other components.Letting a Bazel library under
yapnr/go without//yapnr:package: rules_python then auto-creates emptyyapnr/__init__.pyfiles in its runfiles, and Bazel 7’s sandbox reuse once wrote one through a symlink into the source tree (the file was truncated; six tests failed).Serving the viewer’s static files after
Path.resolve(): in Bazel runfiles every file is a symlink, so a resolved-path containment check rejects all of them. Check containment lexically.Pointing the atopile source index at Bazel runfiles: it skips symlinked
.atofiles on purpose; tests copy the fixture (yapnr.viewer.testing.fixture_copy).Resolving
requirements.lockon linux-x86_64 with the default PyPI index: the torch wheel there needs CUDA libraries the lock does not carry. Resolve on darwin-arm64 or linux-aarch64 (PR6a adds a separate x86_64 lock).Changing imported code or adding allow markers so that the privacy scan passes: the owner chose to refine the scan’s e-mail rule instead (
docs/decisions.md).Running
git filter-repo(orgc, or a branch switch) in the Splanc checkout: experiments run from it. Filter a scratch clone and remove itsoriginremote first (migration plan §7.5).Scrubbing imported machine paths in a later commit: the earlier commits would still publish them. Scrub each snapshot when it is staged, before its commit (PR2).
Reproducing the
board.Removeteardown SIGSEGV on synthetic boards:RemoveandDeleteboth exit 0 there (zones, footprints, tracks, vias, any free order); the evidence is src12i’s run. On H7’s real boardsRemovedid not crash either (it leaks: SWIG’s “no destructor found”).Using SWIG’s “memory leak … no destructor found” messages in H7 logs to find the
Removesites a run reached: only some item types print it (shove’s messages areSHAPE_SEGMENT, notRemove). TraceBOARD.Delete/BOARD.Removecalls instead.Making placement bitwise identical across macOS and Linux (thread count, deterministic algorithms, float64
exp/log): the thread count is already 1, deterministic mode changes nothing, and Adam’saddcmulalso rounds differently between the torch builds (#6).Comparing one seed’s HPWL between placer variants, or across platforms: the per-seed spread (about 100 mm, 5 %) exceeds most effects; compare means over seeds on one platform (#6).