Rust Phase 6 native Cruncher delivery audit

Status: accepted. Phase 5 Plotter-IR is accepted and frozen, and the bounded Phase 6 Windows x64 native Cruncher sequence is closed. The P6_000 Monkey-owned kicad-monkey-native operation process, generated transport, bounded Python client, and Windows platform-wheel lane are accepted. They are transport and packaging foundations, not a promoted Cruncher hard switch. In particular, the repository does not contain a native kicad-cruncher executable or a full Cruncher-native SVG/CLI replacement. The accepted Windows physical-base, compiled-graph, and version-E netlist providers are deliberately bounded exceptions.

Authority and ownership

ADR-011 defines the native Windows kicad-cruncher delivery boundary. The package boundary remains unchanged: kicad_monkey owns low-level source models, strict contracts, native operations, and basic rendering; the separately distributed kicad-cruncher package owns its CLI and artifact orchestration. Monkey must never depend on Cruncher.

The existing public Cruncher distribution is still a Python wheel. Its kicad-cruncher and kcr scripts point to the Python CLI, and it carries an ordinary public kicad-monkey dependency. The Rust workspace also contains the accepted, unpublished kicad-monkey-native P6_000 transport foundation. Its generated envelopes, platform-wheel inclusion, clean-install execution, and independent review are closed evidence. The accepted P6_020 slice now uses that installed operation as Cruncher's selected Windows PCB physical-base serializer without a legacy retry; higher-level and cross-platform provider migration remains staged.

Phase 5 IR boundary is not an SVG or application boundary

Phase 5 closed strict, bounded Plotter-IR producers for footprint, symbol, board, and schematic documents, plus the governed native text-cache path. It froze the producer-specific record and operation contracts. It did not close SVG serialization, Cruncher theming or enrichment, design-review artifact assembly, a native executable, platform packaging, or the public CLI switchover. A Phase 5 document matching Python exactly therefore does not by itself prove that a rendered SVG or a Cruncher output directory matches Python.

The strict Phase 5 contracts remain the transport boundary for native SVG work. Adding a new record or operation arm requires a new contract version or an independently reviewed tolerant envelope; Phase 6 must not mutate the frozen a0 unions merely to simplify application integration.

Bounded delivery sequence

Slice	Required outcome	Promotion evidence
P6_000 native operation transport and package	Add a Monkey-owned, versioned native operation executable or sidecar with bounded byte-oriented input/output, generated-contract decoding, deterministic diagnostics, and fail-before-publication behavior. Package it in an installed Windows artifact without a Rust toolchain or workspace-path dependency. The proving operation may carry a compiled graph and version-E netlist, but it does not yet replace Cruncher design assembly or promote the full design-fact surface owned by P6_030.	Clean-wheel install, protocol-version and malformed-request tests, resource mutations, atomic-output tests, and an installed-process smoke test with the repository unavailable.
P6_010 native SVG	Serialize the frozen Plotter-IR and native text caches to bounded, deterministic SVG. Separate base SVG correctness from Cruncher-owned presentation, theming, links, and document enrichment.	Exact cross-producer SVG vectors, independent semantic SVG goldens, malformed block-state and resource tests, and selected Python-authority geometry/style comparisons. Live KiCad presentation comparison remains provider work because this base profile is not a public-wrapper byte-equivalence claim.
P6_020 no-fallback physical provider	Switch the selected Windows Cruncher PCB physical-base SVG serialization seam to the installed native operation. Python retains PCB loading, Plotter-IR production, filtering, enrichment, theming, virtual/HLR geometry, and orchestration. Missing, incompatible, or invalid native output is an error and never triggers legacy SVG serialization.	Spy tests proving zero legacy-provider calls, absent/bad-sidecar failures, exact typed results, bounded diagnostics, and no partial artifacts.
P6_030 native design facts	Switch the selected Windows Cruncher compiled graph and version-E netlist artifact to one source-bound native result. Python retains Design JSON, netlist JSON, SVG presentation, manifests, and command orchestration. Existing Rust netlist JSON remains test-only and cannot become a public transport until its schema and compatibility policy are explicitly promoted.	Cross-package Python/native fact vectors, hierarchy and occurrence cases, malformed and ceiling mutations, deterministic serialization, and installed-package design-review fixtures.
P6_040 full CLI compatibility	Preserve the public `kicad-cruncher`, `kcr`, and `python -m kicad_cruncher` entry points; command aliases, arguments, output layout, logs, and exit codes; and the primary `design`/`design-review`/`dr` workflow. Any Python facade that remains must orchestrate the native provider rather than provide a fallback implementation.	Installed cross-package CLI tests over success, user error, malformed input, missing dependency, and partial-output cleanup, including exact artifact manifests.
P6_050 exit	Make the selected Windows bot and release artifact use the native path with no fallback, then promote the scope only after every prior slice is independently reviewed. Linux and macOS remain Python unless separately measured and promoted; Windows-first is not a claim of cross-platform hard switchover.	A mandatory Rack exit gate, clean public-distribution install, a content-addressed corpus prerequisite with inherited oracle evidence, package provenance checks, publication of the same hash-bound tested candidates, and a recorded runtime budget.

P6_000 closed on 2026-08-17. Its Rack evidence is L0_062 (15 tests, 0.71 seconds warm) and L1_037 (one real Python-to-Rust operation test, 0.37 seconds warm). The full Rust workspace test suite, generated-contract checks, capability-contained path mutations, and public-sdist to isolated Windows-wheel design-facts execution were also green. Three independent reviews approved the slice. This closure does not promote SVG or any Cruncher fallback switch.

P6_010 closed on 2026-08-17. It adds the generated render-svg operation and bounded plotter-base-a0 serializer over all 30 frozen Phase-5 parity documents: 29 deterministic SVG results and one explicit negative-width safety rejection. Its viewport, fixed base profile, pinned hashes, exclusions, and resource ceilings are described in the native SVG slice. P6_010 itself did not change Python SVG APIs or Cruncher composition. The later P6_020 adapter owns layer views, enrichment, and the selected Windows no-fallback provider without broadening the P6_010 base profile.

Closure evidence was L0_063 (17 tests, 0.19 seconds warm), L1_038 (2 tests, 1.34 seconds warm), and L3_023 (1 test, 11.05 seconds warm; 11.92 seconds with Rack orchestration), plus the green full Rust workspace format/check/Clippy/test gate and three independent approvals.

P6_020 closed on 2026-08-17. It redirects only the Windows PCB physical-base serialization seam, retaining Python PCB loading, Plotter-IR production, layer filtering, enrichment, theming, virtual/HLR geometry, and orchestration. Its bounded request/cache boundary, governed mask-point quantization, no-fallback failures, transactional publication, and installed two-wheel evidence are described in the native physical provider slice. Rack L3_024, package signoff, installed two-wheel execution, affected broad-workflow regressions, and three independent approvals closed the slice. The broader Cruncher hard switch remains staged.

P6_030 closed on 2026-08-18. Its new A1 facts result binds the compiled graph and version-E netlist to the exact current source snapshot and requested source/date/tool metadata. The Windows provider consumes that result once without a Python graph or netlist-writer retry, while Python retains Design JSON and netlist JSON. The bounded protocol, provider, transaction, and evidence boundary is described in the native design facts slice. Rack L3_025, the four-project parity suite, the isolated sdist-built Windows wheel workflow, full native/workspace gates, and three independent approvals closed the slice.

P6_040 closed on 2026-08-18. It treats native CLI delivery as compatibility of the installed Cruncher facade over mandatory packaged Rust providers, not as an unimplemented separately compiled Rust Cruncher executable. It preserves all public command help and the three entry mechanisms, with exact functional acceptance for design/design-review/dr, including arguments, normalized logs, exits, strict manifests, artifact topology, hostile native failures, and transactional cleanup. The bounded contract and evidence are described in the native-backed CLI compatibility slice. Rack L3_026 passed its isolated sdist-built Windows two-wheel gate in 239.86 seconds, composing 116 focused tests; three independent reviews approved the boundary.

P6_050 closed on 2026-08-18. Its Windows x64 exit workflow constructs one hash-bound public-distribution candidate set, requires the content-addressed reviewed corpus ZIP, freezes the accepted native and CLI contracts, and makes release jobs consume the tested Windows Monkey wheel and Cruncher distributions rather than rebuilding equivalents. Corpus semantics remain inherited from the mandatory L3_023 and L3_025 prerequisite gates; the candidate gate binds the ZIP digest and runs installed artifact smokes rather than rerendering that corpus. The exact boundary is described in the Phase 6 Windows release exit. The final promoted-tree candidate passed L3_027 1/1 in 67.38 seconds (68.38 seconds with Rack), and the complete build plus gate took 192.6 seconds wall clock. The refreshed aggregate Rack report passed 7/7 subtests and 43/43 tests, including L99 23/23. Three independent final reviews approved the corrected boundary.

The final hygiene catch-up records 131 exact findings versus the prior 29 after the reviewed Phase 5 and P6_000 through P6_040 Rust expansion. It does not change any policy limit, and future increases remain failures. This debt-inventory refresh is independently reviewable P6_050 governance rather than a claim that the recorded functions were simplified.

Transport and resource rules

Each operation has a named protocol/version, a strict request and result contract, independent byte/count/work ceilings, and bounded diagnostics. Ad hoc internal JSON is not a public contract.
Inputs are read from explicit paths or bytes supplied by the caller. Paths are contained beneath an explicit staging root. The native core does not discover projects, fonts, templates, or output locations.
Outputs are bounded in memory or written to temporary staging and are published only after the complete operation validates. Overflow, truncation, protocol mismatch, missing resources, and malformed native output fail closed.
The Cruncher adapter performs an explicit protocol handshake. It must not retry through Python after native execution begins or succeeds partially.

Packaging constraints

The clean installed-distribution boundary is mandatory. A built Cruncher artifact must continue to declare a normal public kicad-monkey dependency and must not embed repository paths, editable installs, path dependencies, or a copy of the Monkey Rust workspace. A Monkey-owned native sidecar may be shipped by a platform-aware Monkey wheel and discovered by a documented package API. If a separately built Cruncher-native artifact is chosen instead, its source and binary ownership must satisfy the same one-way dependency and provenance rules.

P6_000 adds a platform-tagged Windows Monkey wheel build, archive inclusion checks, an sdist-to-wheel rebuild, and a clean isolated install test for the package-owned executable. Cruncher remains a separate universal wheel and still follows its established Python rendering path. Signing and any separately packaged native Cruncher executable remain later release decisions under ADR-011; public Cargo publication is not implied.

Compatibility and exclusions

The public Python APIs remain compatible through the staged transport delivery. The hard switch happens only for a specifically promoted operation; it is not inferred from the presence of a binary. PyO3 is optional and is not a prerequisite. New WASM exports, browser packaging, downstream pcb_a0, Viz, ALX, and application-specific report formats are outside Phase 6. Cruncher-owned manifests and presentation remain in the Cruncher package even when their physical facts and SVG are native.

Exit decision

All six Phase 6 scope rows now have executable vectors, implementation evidence, semantic/resource suites, and Rack owners. The selected Windows public install completes design/design-review/dr without a Python physical fallback, and CI and release use the same tested candidate set. Phase 6 is therefore accepted for Windows x64. Linux and macOS retain their established Python providers, and Python continues to own Cruncher CLI parsing, orchestration, presentation, enrichment, and transactional publication.