Skip to content

Human action 18

Jerry Holland edited this page Sep 29, 2026 · 4 revisions

Bounded native read-engine decision accepted

Canonical decision: #18. The maintainer accepted ADR-0027. Implementation is tracked separately in #83.

Decision and scope

The first phase-1 increment covers five existing resolution-catalog read operations through an optional Rust binding, initially for files-backend SHA-1 repositories. Domain validation and mutations stay in their current paths; unsupported operations/profiles and unavailable bindings use Git fallback.

The accepted Gate A experiment uses the matching identified Windows host lab-windows-a, the reduced resolution-catalog workload, and the published Git data-acquisition reference. The limit is 110%, with no additional allowance. Node startup is excluded once; initialization and validation remain in the phase and whole CLI time is reported separately. Native execution and byte-identical domain behavior must be proved independently from the latency ratio.

Execution and verification

Owner role: The implementation maintainer for #83; the repository maintainer for any change to the accepted scope or budget. No further owner scope approval is needed merely to start the required dependency review.

  1. Read ADR-0027 and the accepted decision on #18. Preserve its five-operation boundary and supported-profile/fallback rules.
  2. Before native code, record and review exact dependency versions, features, supported capabilities, licensing/attribution, safe-core/FFI controls, and Windows/Ubuntu toolchain and optional-prebuild support on #83. Confirm that the package still works with no binding. Escalate only actual changes to the accepted scope or costs.
  3. Implement and validate under #83. Require actual native execution, all six functional modes on Windows/Ubuntu plus Node 20, Git-output equality, bounded hostile-input handling, zero Git processes in the supported catalog phase, and working fallback. A binding-missing run is only seam qualification.
  4. Before shipping, perform the ADR's three independent quiet matching-host comparisons, at least nine samples each, alternating timing order. Every comparison must meet 110%. Preserve raw samples and public-safe build/host evidence on #83; do not change the denominator or baseline to pass.
  5. If the bounded attempt misses, retain Git as the delivered engine and record the result. Further tuning or scope expansion follows the explicit decision rule in ADR-0027. If it ships, record the first version and the two-minor-release sunset deadline. Verify read-only behavior, fsck, worker cleanup, and uninstall.

The decision no longer waits on OneDrive. #14 remains its own qualification task; #19/#41 still require POSIX participation and sustained workflow evidence; #42 retains broader representative budgets. Wider phase-1 work and phases 2–6 do not open from this increment. Windows-only release latency does not waive POSIX functional support or Gate B.

Completion of #18 means the accepted decision is delivered consistently in ADR, product/testing guidance, issue, and board. It does not claim a working backend. Retain this procedure and its Human TODO link as the record of the resolved owner action. Implementation results belong on #83.

Current build and usage instructions are in Optional native resolution reads.

Back to Human TODO

Clone this wiki locally