Domain vocabulary for comment-checker: words with codebase-specific meaning that solutions docs and instructions cite without redefinition.
A comment that says, in prose, what its adjacent code already says — the
filled-in words or the verb-to-operator mapping ("increment" ↔ +=) are
the same fact twice. Distinct from a justification, which adds a reason the
code cannot show.
A comment that names the control construct the adjacent code already
displays ("loop over each" beside a for) — the kind that reads like
spoken code and adds nothing.
The tokens or verb→operator matches a restatement verdict was built on, published with the flag so the reason is checkable against the code. A verdict that cannot cite verified evidence is not emitted.
Structural context (the adjacent-code window) that cannot vouch for the comment because the input was a fragment: Edit/MultiEdit boundaries truncate what the hook can see. Verdicts that depend on context never convict on unreliable context — they fail open (spare). Explicit text-only rules (unlike restatement) still apply on fragments, because they need no context.
The production parse's adjacent-code snippet when the hook had the whole file and the code window is trustworthy; the only input context-aware verdicts may cite.
A per-kind/precision-recall gate on the corpus that trips when a kind's classifier weakens — including a single-case kind that goes wrong — so a weakness in one kind cannot hide inside an aggregate F1 score.
The root npm package (@systemfsoftware/claude-code-comment-checker) whose
bin is the comment-checker shim. It resolves the host platform package by
identity at runtime and spawns the binary — the only package that declares a
bin.
One per os-cpu pair (-linux-x64, -darwin-arm64, …), generated from
scripts/lib/targets.json: ships only the compiled binary and its
manifest (os/cpu/libc fields, no bin). The launcher's
optionalDependencies pins all five to the release version.
The committed launcher manifest never lists these packages as
optionalDependencies — pnpm cannot lock unpublished platform packages
(pnpm#3960), so the pins are injected from the targets table at publish
time; absence in-tree is expected, not a defect.
A matrix row in the release workflow: one platform/arch build, gate, smoke, and publish run on its native runner. Platforms publish before the launcher, and the release cannot proceed if any lane fails.
The turbo task cache that stores the mutation gate's result keyed on the gate's input set; unchanged inputs replay the stored verdict instead of re-running the mutant loop.
Two hash surfaces govern it. The task hash is the gate's own key: every file that can change the verdict must be an explicit task input — automatic coverage extends only to the script body and declared env values. The transport key (the CI cache step's key) must be a superset of the task hash surface, because an exact-key restore suppresses the post-run save: a narrower transport key does not replay stale verdicts, it discards fresh ones.
An index that lets a Claude Code / OMP installer discover and pull a plugin from a marketplace source. Distinct from the plugin manifest: the catalog declares the marketplace and lists each plugin with a source, while the manifest carries the plugin's authoritative name and version. A catalog carries no version of its own — the version lives only in the manifest, so the catalog is not a release surface.
The explicit, enumerated set of manifests a release bump updates atomically. A manifest not on the list is intentionally outside the gate: keeping a file off the surface means its version can never drift out of step, and adding a manifest to it is a deliberate decision, not a default.
A verification mechanism (a hook, a gate, a check) exits success while checking nothing — the failure mode of going green. Distinct from a false negative, which is a wrong verdict on a real check; a silent pass means the check never ran, the input never reached it, or its failure was swallowed. A hook must never claim a check that did not happen: absent checker exits with guidance, a flagged write exits with the report, and no surface ends in a blanket success. Avoid: fail-open (reserved for the classifier's deliberate spare)
- "context" had been used for both the language (scope/position) and the evidence (adjacent code) — these are distinct; adjacent syntax is the only context that carries the mention.
- "fail open" in the classifier spares a comment it cannot judge (a deliberate conservative verdict) while "silent pass" in the hook gate exits green while checking nothing (a defect) — distinct meanings, distinct domains.