perf(hooks): batch pre-commit staged-lock auto-claim into one call

[NagyVikt]

Summary

• perf(hooks): batch pre-commit staged-lock auto-claim into one call

Test plan

• verified locally

[NagyVikt]

Decision: APPROVE — 0 CRITICAL, 0 HIGH, 0 MEDIUM, 2 LOW.

Collapses the pre-commit staged-lock auto-claim from one gx locks claim process per staged path into a single batched call, mirroring the shape the staged-deletes branch below already uses. 15012ms → 1327ms on a 30-file commit across 12 worktrees (11×).

The fallback is the part worth calling out: locks claim is atomic, so one foreign-owned path rejects the whole batch and claims nothing — which would leave locks validate --staged reporting every staged file as unclaimed instead of the one that actually conflicts. Falling back to the per-file loop on batch failure keeps the diagnostics exact, and incidentally makes an ARG_MAX overflow degrade gracefully rather than silently skipping claims.

Paths exercised during review:

Path	Result
Delete-only commit (empty non-delete set — the set -u empty-array trap)	Hook exits 0, no expansion error
Foreign-owned path in the batch	Blocks; names only the conflicting path; others claimed by fallback — output identical to pre-change
Variable collisions / downstream consumers	None; _staged_deletes block independent
set -e vs if ! cmd and [[ ... ]] && continue	Both exempt per bash rules; confirmed empirically

LOW-1: conflict case now costs N+1 spawns (only on a path that ends in a blocked commit anyway — correct trade).

LOW-2 (pre-existing, not introduced here): --name-only returns core.quotePath-escaped paths for non-ASCII/spaces, which silently fail to claim. The replaced loop read from the identical source, so exposure is unchanged. Proper fix is -z + read -r -d '' across both this loop and the deletes loop — out of scope for a perf change.

Tests: 41 failing on branch, 41 on main, sets byte-identical — zero new failures against a baseline-red suite. Lint skipped (biome absent locally; its scope excludes templates/).