History import manifest¶

Sources¶

The snapshots are read-only sources; each was copied to a scratch directory, checked against the digest below, scrubbed there and then committed. The engine digest is the sha256 of the sorted path<TAB>sha256 lines of the engine paths (below), without __pycache__. The hierarchical snapshots carry F217’s hashes.json unchanged, which does not describe them, so the digests are used instead.

• F217 is the fresh217 source freeze (2026-09-25) of Splanc’s working tree: the Codex agent’s uncommitted engine work. F222, the fresh222 freeze, adds the Electrical221 integration; its hardware/pnr is byte-identical to Splanc’s working tree, which has not changed since.

• The author date of a snapshot commit is the capture time (F217, F222) or the newest source modification time; times are Pacific (UTC-7). The committer date is the time of the import.

• Directories that are engine-identical to one above produce no commit: src9.frozen and src10.base (= src8b), src10b and src11.base (= src10), src11.frozen and src12.base (= src11), src12n.base (= src12h), src14.base (= src12i), src15.base (= src14), src13.base (= src12b) and src13 (= src13.r8). Earlier captures of the src13 line are older versions of src13.r8 and are not imported. The snapshots src2 to src8, src8.frozen and src9 no longer exist; src8b carries their work.

• src15 was re-verified against its digest immediately before its commit: it is the engine of a running experiment and was edited until 03:19 on 2026-09-30.

Engine paths¶

hardware/pnr/                          engine package, tests, testdata, regression, si_models,
                                       BUILD.bazel, pnr.bzl, WORKLOG.md
hardware/tools/keyhole_region.py       the only tools the source freezes carry
hardware/tools/freeze_mini_inputs.py
hardware/tools/audit_pair_contacts.py  Splanc working tree only (untracked there, see PR1 above)

Not imported: the Splanc board design (hardware/splanc_dev outside the engine’s own test data, atopile sources, constraints, sourcing), the snapshots’ module and Bazel configuration and patches, hashes.json and metadata.json, the snapshot-level worklogs (four of them hold machine paths), the merge and SI evidence directories, __pycache__ and every run directory. The Splanc working tree’s viewer changes (pnr_live) go to PR4 (decisions). No step deletes an engine file; every step adds or modifies files only.

Lineage and commits¶

The verified lineage differs from the plan’s first version (§7.3 is updated in this pull request): src10.frozen precedes src10 (src10 adds the owner’s 0.15 mm stretch floor); src14 is a lineage step; src13 (USB pairs) and src12n (N-0001) are side lines that join at src15; and Electrical221 is in no hierarchical snapshot: it forks from F217, like the whole hierarchical line.

main
 └ F217 ─┬─ src8b ─ src10.frozen ─ src10 ─ src11 ─ src12 ─┬─ src12h ─┬─ src12i ─ src14 ─┐
         │                                                │          └─ src12n ────────┤
         │                                                └─ src12b ─ src13.r8 ────────┤
         │                                                           merge = src15.r1 ─┘
         │                                                           src15
         └─ Electrical221 (F222) ──────────────────────────────────  merge Electrical221
                                                                     Bazel adaptation
                                                                     docs

In all, 228 files change against main (+72,122 / -614 lines) in 350 new blobs; the largest is 327,031 bytes (testdata/feedback/board-placed.json), and none exceeds 600 KB.

• Snapshot commits. Each snapshot commit’s engine tree equals its snapshot for every file, except the four scrubbed files below; checked for each commit against the snapshot directory. The octopus merge records the three lines as parents and has the tree of src15.r1, the validated result of the per-file 3-way merges done in the experiment area.

• Electrical221 merge. git merge of the Electrical221 commit into src15 conflicts in two files, resolved as follows. In pnr/via_coalesce.py, the report’s summary keeps src15’s removed_cycle_tracks and adds Electrical221’s reconstructed_cycle_tracks and removed_cycle_length_mm. In pnr.bzl (two hunks), Electrical221’s PNR_FULL_ELECTRICAL_POOL branch uses src15’s headless "$_KI_CLI" in both branches, and the action tools are the placer, the native and electrical controllers, and the SI tool when si_report is set. BUILD.bazel, pnr/track_graph.py and tests/test_via_coalesce.py merge cleanly.

• Build. Every commit before the adaptation carries Splanc’s BUILD.bazel (it loads @pypi), as PR1’s import commits do: only the adaptation commit and the docs commit build in yapnr, so a bisection inside this series needs the adaptation applied.

Content scrub¶

Before the scrub, tools/privacy_scan.py over every distinct engine blob of the sources (564 blobs across all snapshots and the working tree) reports 37 findings, all absolute volume paths, in four files. A targeted search of the same blobs for host and network names, network addresses, the login name, home and scratch paths and personal addresses finds nothing else. The four files are scrubbed when each snapshot is staged, before its commit, so no commit carries the paths:

The fixtures’ findings are provenance records (source.path); the tests’ are the path of a recorded block case and of an experiment virtualenv. The replacement patterns stay in the operator’s scratch directory. No other file differs from its snapshot.

Privacy and tree checks¶

Authors and committers are Claude Agent with the owner’s public commit address; the author dates are those in the source table. Every message ends with the session’s two trailers. CI’s history scan has included merge diffs (git log -p --diff-merges=separate) since the engine hygiene change, so they are now scanned in CI too.

Bazel adaptation and pass set¶

The adaptation commit re-applies PR1’s adaptation to Splanc’s newer BUILD.bazel:

• the @yapnr_pypi load;

• the SI test comprehension listed test_si_annotations.py and test_si_extract.py twice for their own targets, which Bazel rejects as a duplicate label, so the whole package failed to load; each source is now listed once;

• si_integration_test depends on :pnr_fab: its RouteMainTest cases import pnr.route.__main__, which :pnr_route excludes;

• PR1’s manual tags, re-applied with an updated comment (below);

• CI runner fixes found on the PR’s first run: placement_test is size medium (over the 60 s small timeout on the 4-vCPU macOS runner), and test_si_runner’s nice check is relative to the test process (GitHub’s macOS runners start jobs at nice -10, so the absolute >= 5 check failed although the +5 increment works). That test file is the only non-BUILD edit in PR2; it is yapnr’s own SI test, not Splanc-imported history.

Pass set, from bazel test //... --config=ci with --config=lowmem and --nocache_test_results on the development Mac (hermetic Python 3.11, torch 2.3.1, numpy 1.26.4):

• 100 of 100 test targets pass: 92 imported (the 69 that passed at PR1 and 23 newly wired: bus_classes, cost_capture, cost_inspect, cost_outline, cost_pool_context, probe_cost_capture, electrical_pool, fanout_reserve, group_backtrack, movable_groups, partial_cycle, relative_rows, nine si_* targets, si_integration and terminal_min_width) and yapnr’s 8.

• 85 test methods in 17 imported targets skip: they need KiCad’s pcbnew, the headless KiCad copy, or the opt-in live SI cases (PNR_SI_LIVE=1). The KiCad lane (PR6a) runs them.

• Still manual, each with a comment in BUILD.bazel and listed in the plan’s appendix B:

  • placement_budget_test, progress_budget_test, retry_controller_test: they fail earlier than at PR1 with FileNotFoundError: 'fixture-python'. They patch native_loop.subprocess.run, but since src8b the native loop starts its workers through pnr.proc.run, so the fake interpreter of the recorded fixture is executed. The engine code is unchanged in this PR; the tests are fixed with the fixtures in PR3.

  • detail_route_test: unchanged, 848 s on the development Mac and test_drc_clean_by_construction fails as at PR1.

  • orientation_test: size medium and manual as at PR1 (Studio-Fug/yapnr#6; it passes on macOS). Re-enabled after PR2, in the engine hygiene change: placement is deterministic per platform only, so its HPWL check now compares the three-seed mean with a 5 % margin, and a synthetic board checks that the search finds known best angles (decisions). Still medium.

• pnr.bzl stays unloaded, and hardware/tools/BUILD.bazel is unchanged: keyhole_region.py is still the only tool a test needs.

Test wiring¶

At the head, 68 test files under hardware/pnr are not wired to a Bazel target: the 25 of PR1 and 43 new ones. tools/check_test_wiring.py --scope hardware/ reports 73 there (77 under hardware/ with the four outside hardware/pnr that PR1 listed): it looks for each file’s name in the BUILD file, and five SI test files (test_si_deck, test_si_metrics, test_si_models, test_si_physics, test_si_runner) are named only through the si_%s_test list comprehension, which wires them. They stay unwired until PR3 moves the tests under tests/.

Known leftovers for PR3¶

No engine code changes in this PR; these are carried forward (plan appendix B lists the defects). Struck items were done after PR2, in the engine hygiene change (branch claude/engine-hygiene):

• Workers without a time bound (AGENTS.md: a subprocess without a timeout is a bug). New in PR2: pnr/electrical_pool.py (the pool workers), pnr/hier/native_block.py (two calls), pnr/mc/halving.py (one), pnr/transaction_cleanup.py (one), the parallel pair trials in pnr/paired_bootstrap.py (polled until a stop event, no deadline) and the cleanup pass of hardware/tools/keyhole_region.py. From PR1’s code, still unbounded: drc_warm, electrical_repair, full_iteration, geometry_optimize, paired_bootstrap, plane_leaf_repair, power_detour_repair, staged_signal and regression/run.py. Two of PR1’s are fixed by src8b (native_loop, via_coalesce through pnr.proc.run). All move to pnr.proc.run. Done: every one runs under a pnr.proc deadline (PNR_WORKER_TIMEOUT, PNR_PHASE_TIMEOUT or PNR_EVALUATION_TIMEOUT, decisions); a timeout kills the child’s whole process tree, so the workers of a killed evaluation go with it, and //hardware/pnr:proc_test fails on a new unbounded call. Left by design: the opt-in warm DRC host daemon of drc_warm/launch_host.py (ended by its session).

• Engine code tied to the Splanc design. New in PR2: pnr/feedback/signals.py recognizes annotation sources by a hardware/splanc_dev/elec/src/*.ato pattern, and pnr/hier/native_block.py and pnr/mc/halving.py default to hardware/splanc_dev, its .ato source and its fab JSON files. From PR1: the defaults of full_iteration, geometric_native, drc_warm and freeze_mini_inputs.py. These become project inputs (PR3c, PR3d).

• Tests that read Splanc design files (hardware/splanc_dev, the experiment area, or parents[4]), skipping or failing without them: power_topology_golden.py, test_fab_profile, test_feedback_moves, test_feedback_signals, test_halving_generations, test_macro_hull, test_place_pair_weights, test_shove, test_shove_native, test_si_extract, test_si_integration, test_si_live, test_src2_library, test_synth_native_generations, test_terminal_min_width and test_via_in_pad (PR1: test_ingest). They get fixtures from the examples or become manual tests of an external project (PR3c, PR6b).

• Splanc design excerpts in the engine’s test data and models. None is private, but they are Splanc Mini data: testdata/feedback (25 JSON files of one converter block, the largest 327 KB), testdata/power_topology (provenance names the Splanc board), testdata/si (an excerpt of one routed board, three comment lines of .ato, and SOURCE.json, which names the run it came from), si_models/parts.json (Splanc part directories and supplier part numbers) and si_models/profile-ws2812b-din.json (a document path in the experiment area). Regenerate from the examples (PR3c, PR6b). PR1’s testdata/splanc_dev board is the same case; it also carries mode 100755 from PR1.

• Install paths. Path.home()/'Applications/KiCad-headless.app' defaults in si/extract, si/models, si/runner and regression/{run.py,README.md}; the output/pnr-regression-runtime/bin/python default in regression/run.py; the /Applications/KiCad defaults in 12 modules, regression/{run.py,README.md} and pnr.bzl: toolchain discovery (PR6a, PR3c).

• GUI-bound tool: hardware/tools/audit_pair_contacts.py starts a wx.App (outside the loop): make it headless or drop it (PR3b).

• Default-on behaviour: Electrical221’s via_coalesce/track_graph changes (partial-cycle cleanup, barrel-contact bridges) are unflagged and now yapnr’s default; the H7 run uses src15 without them. Gate or A/B them (decisions). Gated: PNR_PARTIAL_CYCLE_CLEANUP=1 and PNR_BARREL_CONTACT_BRIDGES=1, both default off, so the default is src15 again; the A/B that would turn them on is still open.

• Unchanged from the plan: the board.Remove audit (PR2e) (done: 27 calls now use board.Delete, 3 stay, //hardware/pnr:board_delete_test guards each call; replayed on the H7 boards for five sites with identical results, a full rung-1 evaluation with it is still open) and the missing pnr.capacitor_intent (PR4b).

Reproducing the series¶

With the variables of plan §0 bound locally and the scrub patterns in $SCRATCH, for each snapshot in the order of the commit table:

rsync -a --exclude __pycache__ "$SNAPSHOT/hardware/pnr/" "$SCRATCH/snap/$NAME/hardware/pnr/"
cp "$SNAPSHOT"/hardware/tools/{keyhole_region,freeze_mini_inputs}.py "$SCRATCH/snap/$NAME/hardware/tools/"
# compare the engine digest with the source table, apply the scrub, then:
rsync -a --delete --exclude __pycache__ "$SCRATCH/snap/$NAME/hardware/pnr/" hardware/pnr/
git add -A hardware/pnr hardware/tools && GIT_AUTHOR_DATE="$DATE" git commit -F "$MESSAGE"

The octopus merge is written with git commit-tree from the src15.r1 tree and the three parents; the Electrical221 merge is a plain git merge with the two resolutions above.