Hook gates are recorded rather than registered decisions

Context and Problem Statement

decisions/registry.json gives every control-flow decision one id, one program and one home, and scripts/decision-check.py’s R10 (#252) refuses any tracked script that is neither a registered decision’s program.path nor a key of not_decisions. Widening R10’s enumeration to hooks/ (#307) forced the question the narrow globs had let the kit avoid: a PreToolUse hook looks exactly like a decision. hooks/roseline-gate.sh branches on ROSELINE_GATE=off|on, then on a launcher probe, then on the file’s own extension and project context, and emits allow or deny — #307’s own Problem section calls it “precisely the shape decisions/registry.json exists to govern” and its Alternatives section rejects recording it as “a false claim in the registry”.

Two facts pull the other way, and both were discovered only by reading the hook end to end. A registry row is a claim about a verdict: verdict.source names how scripts/decide.sh extracts one, and a hook has none to extract — its deny is a permissionDecision on stdout in Claude Code’s PreToolUse envelope, consumed by the harness, never by the dispatcher. And R7 requires an owner document that invokes the decision inside a fenced block; a hook is invoked by hooks/hooks.json matching a tool name, so there is no owner to name and nothing for R7 to check.

Considered Options

• Register each hook, inventing a verdict.source for the PreToolUse envelope and relaxing R7 for decisions with no owner document.
• Record each hook in not_decisions, with a reason that says which of the two tests it fails.
• Leave hooks/ outside R10’s enumeration, as before #307.

Decision Outcome

Hooks are recorded in not_decisions, not registered: a registry row is a claim that decide.sh dispatches this program and that an owner document invokes it, and a hook satisfies neither — so a row would have to be paid for by weakening R7 and by adding a verdict source no dispatcher implements, buying a claim about a program CI would still never run. #353 already recorded hooks/git-write-gate.sh this way; #307 records hooks/roseline-gate.sh on the same terms, and R10 now enumerates hooks/ so a third hook must be classified rather than ignored.

Consequences

The registry’s own guarantee gets narrower and more honest: it governs the decisions decide.sh runs, not every allow/deny in the kit. What keeps that from being a hiding place is R10 — since #307 no hook can be silently absent, and each one’s not_decisions reason has to name why it is not a decision, so the classification is a per-file claim a reviewer can dispute rather than an omission nobody sees.

The counter-argument is real and is the reason this record exists: a fail-open access-control gate is exactly the kind of file the decisions/ system was built to catch, and “recorded as not a decision” reads like the drift the system prevents. It is not the same thing — a recorded file is enumerated, named and reasoned about — but the resemblance is close enough that #307 argued the opposite position in writing before the hooks were read. Reverse this only by giving decide.sh a PreToolUse verdict source and R7 an answer for an ownerless decision; both hooks then move together, because ADR 0002 already establishes they take the same terms.