fix(gate-22): one missing _note emitted THREE findings — a count is not a defect count

[rubenvdlinde]

Follows the finding-vs-defect thread from #238 and #228.

Cause

Ajv with allErrors: true reports the applicator keyword alongside the leaf cause. The manifest schema states its _note rule as a nested if/then/else (nextcloud-vue#315 relaxed it so a Cn[A-Z]\w+ component is self-documenting), so one missing _note on one page emitted three lines:

at /pages/0: must have required property '_note' (keyword=required)  ← real
at /pages/0: must match "else" schema (keyword=if)                   ← echo
at /pages/0: must match "then" schema (keyword=if)                   ← echo

That is how a gate run reported "240 violations" for roughly 132 defects: 51 missing _note entries contributed 153 lines between them. The merged AppHost schema also asserted additionalProperties twice over the same node, emitting an identical line twice at /.

An inflated count is not cosmetic. It is what makes a repo's manifest debt look insurmountable and stops people reading the log at all — the same reason gate-46's summary now prints distinct targets alongside raw findings.

Measured

Fixture: two type: custom pages, each missing _note.

	findings	exit
before	10	1
after	5	1

Same five real defects. The verdict does not change — only the line count.

Collapsing errors is deleting evidence unless it is provably lossless

That is the whole risk here, so each property has its own test:

• an applicator error is dropped only when a concrete sibling exists at the same instancePath — a lone if failure really is the only signal at its path and still surfaces;
• dedupe is by (path, keyword, message), not by path — three distinct defects at / all survive;
• no non-empty error list can collapse to zero.

Blast radius

The five re-measured repos (portaliq, doriath, openconnector, larpingapp, procest) are all clean on gate-22 today, as is shillinq — its 240-violation run has since been burned down. So this changes no current verdict; it stops the next repo's number from being inflated threefold.

Tests

New suite test_check_manifest_ajv_collapse.js, auto-discovered by run-helper-suites.sh (27 → 28 suites, 0 failed). It is unit-covered without Ajv, runs end-to-end when Ajv resolves, and refuses to run rather than pass vacuously if the export is missing.

check_manifest.js now exports collapseAjvErrors when required as a module (require.main === module still runs main() as a script), so the helper is testable without Ajv and without validating a manifest as a side effect.

Four mutations:

mutation	assertions that fail
collapse disabled (pre-fix behaviour)	4
applicator dropped unconditionally (lossy)	3
dedupe by path only	2
collapse returns [] (blind)	9

One methodology note worth passing on

My first M3 run reported FAILs=0 — an apparently surviving mutant. It had not survived: the mutation never applied, because my match string was wrong. A mutation that silently fails to apply is indistinguishable from a surviving mutant, and it reads as reassuring rather than alarming. Every mutation in the final run asserts old in source before writing, and the re-run killed all four.

The same check also caught two stray NUL bytes my own edit had introduced into the dedupe key — repaired at byte level, with an assertion that exactly two bytes changed. Both were caught before merge.