Batch out-of-process CheckLibrary runs (collection-scoped)

[abegong]

Background

The CheckLibrary spec and plan introduce CheckLibrary as the abstraction behind every check type, with a SchemaLibrary capability and an out-of-process invocation seam (Vale being the motivating example).

The first cut runs an out-of-process library per item: the engine launches the external process once per file. That is the simplest correct path and matches how every per-item check works today, but it pays process-startup cost on every file. Over a large collection that is wasteful.

What to build

Run an out-of-process library once per collection instead of once per item:

• Gather every item in the collection, invoke the external tool a single time over all of them (e.g. vale --output JSON <files>), and map each finding back to its file via Violation.File.
• Reuse the existing collection-scoped check pass (checks.CollectionCheck / RunCollection, engine.collectionChecksFor, the whole-collection re-scan in cmd/check.go) rather than inventing a new lifecycle.
• Accept that a single-item selector (katalyst check notes/foo) triggers a whole-collection run, which already holds for collection-scoped checks like filesystem_unique_filename.

Notes

• Depends on the CheckLibrary abstraction landing first, and realistically on the Vale library existing as the first out-of-process consumer.
• Spec Open Question 5 records the per-item-first / batch-later decision and the tradeoff.

🤖 Generated with Claude Code