The local Rust library and lesscode-core CLI now prepare and compare supplied
Python and TypeScript capabilities. They measure complete replacements and admit
source-bound behavioral evidence. The core does not propose edits or execute
project tests. The existing discovery skill supplies the hypothesis and reviewed
scope; maintainers decide adoption. Analysis needs no hosted account, commercial
analyzer, Python/Node process, MCP service, solver or accelerator hardware.
Native correctness passed on Windows (102 tests), Linux (103) and macOS (103), with zero skips. The previously missing platform evidence is now supplied by Actions run 41333477507. The final gate records the resulting scoped feature GO. This does not establish general compression efficacy or whole-project simplification.
Build with the prerequisites in native setup, then run from the repository root. Use a fresh sibling output path on each invocation:
cargo +1.98.1 build --locked --manifest-path native/Cargo.toml
$pair = 'docs/experiments/native-replay/ats-run/recorded'
native/target/debug/lesscode-core.exe compare --frozen "$pair/frozen.json" --baseline "$pair/baseline" --candidate "$pair/candidate" --evidence "$pair/candidate-evidence.json" --artifact-root $pair --out D:/Source/lesscode-first-native-result
Get-Content D:/Source/lesscode-first-native-result/report.mdThis readmits retained ATS observations; it does not rerun them. The fresh
readmission receipt records the
actual external output, accepted result and equal specimen hashes. Retained
JSON and
Markdown contain the result.
Use the discovery recipe to freeze a new manifest,
run unchanged checks externally on both copies and supply their artifacts.
prepare --manifest MANIFEST --baseline ROOT --out FRESH.json freezes the trial;
compare requires that freeze, both roots, evidence and explicit artifact roots.
Reports require new output outside the source roots. Exit codes are 0 for
prepared/accepted, 1 for rejected/not-demonstrated, 2 for incomplete and 3 for
invalid/baseline-ineligible/I/O failure. See CLI help receipt.
| Complete owned replacement | Native physical lines | Native statements | Outcome/adoption |
|---|---|---|---|
| ATS Python reporting | 815 → 634 | 345 → 184 | Accepted; no upstream adoption |
| NPU Python provider/helper | 799 → 729 | 264 → 246 | Accepted; no upstream adoption |
| TypeScript degradation capability | 400 → 381 | 93 → 82 | Accepted; no upstream adoption |
| Lesscode Python contracts | 2342 → 2310 | 1599 → 1570 | Independently reviewed; exact policy-contract replacement adopted |
These are separate, manually selected scopes and targets. Native versioned statement/decision definitions differ from historical Radon/Lizard measures. Call layers remain unavailable where unresolved; reviewing ownership does not resolve dynamic dispatch. ATS/NPU task 6.1 freezes with stricter call-layer targets remain incomplete; later source-target trials are separately frozen results. Known behavioral mutants, moved helpers, narrowed/reclassified scope, missing checks and stale evidence are rejected or incomplete in the retained trials and native fixtures. Failed preparation/candidate attempts remain in their records. No pooled success rate, repository-wide shrinkage, maintenance savings, runtime speedup or hardware result follows from these counts.
Fresh Windows checks on 2026-11-01 passed: locked nextest 0.9.143 ran 102 tests,
102 passed with no skips; doctests ran zero tests successfully; build, rustfmt
and clippy with warnings denied passed. The built executable reached usable
--help and accepted the real ATS pair. Raw test log,
doctest log,
build log,
fmt log and
clippy log retain command outcomes
with sibling exit-code files. Parallel test instructions retain
the separate doctest gate.
The self-trial completed after external fixtures. Its frozen isolated baseline and candidate each collected the same 515 Python identities: 504 passed and 11 existing Windows/platform skips. Both passed 171 relevant checks and seven independent methods with adverse controls. Explicit root/config/pythonpath and source guards in every worker establish which implementation ran. The trial record links commands, JUnit, failed earlier runs, source guards and provenance; collector exit zero alone was never sufficient. Fresh live contract checks after adoption passed 70 tests and 104 subtests. The byte comparison receipt verifies all 96 live source/test files equal the admitted isolated candidate; generated egg-info and bytecode are excluded explicitly. The full 515-test suite was not repeated in the live checkout. Approved parallel-tooling/config changes are separately enumerated in integrity.json.
Self-workflow execution is complete for the selected supported Python contracts
capability; demonstrated self-compression is 29 statements, 32 physical lines,
17 decisions and three functions removed. The target was at least 25 statements
with no line/decision/function increase. Only the independently approved file
with SHA-256 65e939cc9e1d3acac3b0aed289d0283e9a160f28875a5f5a0d79cdb509a8dcfd
was adopted. Rust self-analysis and TypeScript self-compression remain untested;
there is no whole-project maturity claim. A rejected beneficialness hypothesis
would retain baseline; no deletion quota is required. Missing evidence would keep
the corresponding maturity claim open.
Earlier WSL foundation checks and the subsequent SIGBUS attempt remain historical records. Current local Dagger validation and the hosted three-platform run passed formatting, Clippy, nextest, the separate doctest command (zero doctests), build and CLI help. Linux runs in Dagger; Windows/macOS execute the same script on native hosts. The durable audit summary records the tested merge revision, identical native source tree and artifact identities; raw artifacts have 14-day retention.
The footprint receipt counts 6,149 physical Rust source lines including inline unit tests, separately from 973 integration-test lines. It cannot provide a production-only Rust total; Rust parsing is outside the supported language scope. The new readiness verifier is 69 research-support lines. The self-trial separately records 151 automated check lines and 592 research-support lines in its footprint, with production/test/research categories in whole accounting. These are implementation costs, not specimen savings. Locks, dependency notices, fixtures and reports add further maintained material. Counts do not establish net repository reduction or effort saved.
Hashes establish consistency, not trusted execution or hostile-writer isolation. Finite checks cover declared domains; external services, arbitrary runtime values and universal equivalence are unproved. Source review, dependency knowledge and adoption remain human/agent responsibilities. The native dependency direction matches the design: model; scope/metrics/compare; library orchestration; thin CLI. The retained Python CLI and unfinished rule-engine work are separate from native analysis.
The 32-criterion coverage map and readiness gate record the precise validation boundary.