fix(BUG-023): rewrite claude-mem hook probe to materialize+break (cross-OS)

[mlorentedev]

Summary

BUG-022 (PR #87) appended head -n1 after the while-loop in the BUG-015 hook-resolution probe, but the form still raced under set -euo pipefail when 2+ candidates matched in ~/.claude/plugins/cache/thedotmack/claude-mem/: the consumer closed after the first line, leftover printfs in the while loop got EPIPE, pipefail propagated 141 to the $(...) and set -e killed healthcheck.sh mid-section 4/12. setup-linux.sh then logged a false-positive WARNING: healthcheck reported one or more FAIL items.

Race-free pattern: materialize candidates into a variable first (no pipe), then iterate in pure bash with break — no consumer-close, no EPIPE, no pipefail propagation. Supersedes BUG-022.

Cross-OS parity: healthcheck.ps1's bash sub-invocation never observed the symptom (its bash -c context doesn't set pipefail), but the race-free form keeps both probes structurally identical. Applies the recently-captured lesson "the detection probe MUST use the race-free pattern, not faithfully reproduce broken upstream behaviour".

Empirical evidence

This user has 6 versions of claude-mem in cache (12.7.4 → 13.3.0), so the race fires deterministically on every setup run. Post-fix healthcheck.sh completes all 12/12 sections; probe resolves to 13.3.0.

$ ls ~/.claude/plugins/cache/thedotmack/claude-mem/
12.7.4  13.0.0  13.0.1  13.1.0  13.2.0  13.3.0

# Before:
$ ./scripts/healthcheck.sh; echo "EXIT=$?"
...
[4/12] Key Symlinks
  PASS: claude-mem@thedotmack installed (BUG-014 canonical check)
EXIT=141                                       # SIGPIPE; sections 5-12 never ran

# After:
$ ./scripts/healthcheck.sh | grep BUG-015
  PASS: claude-mem hook path resolves to: /home/manu/.claude/plugins/cache/thedotmack/claude-mem/13.3.0 (BUG-015)
# Healthcheck now completes all 12/12 sections.

SDD skip rationale

Atomic fix to an existing safety-net probe; production diff ~42 LOC (under the 50 LOC spec-gate threshold). Same shape + precedent as PR #87 (BUG-022) and PR #79 (WIN-001b). No public-contract surface touched (probe is an internal helper inside healthcheck.{sh,ps1}). Lesson capture goes to vault post-merge.

Test plan

• shellcheck scripts/healthcheck.sh clean
• bash -n scripts/healthcheck.sh clean
• bats tests/setup-linux.bats --filter "BUG-023|BUG-015" → 3/3 PASS
• bats tests/ → 765/765 PASS, 0 fail (no regressions)
• Live healthcheck.sh run completes all 12/12 sections, exit no longer 141
• Probe resolves to expected path (13.3.0, latest cache version)
• 2 new parity asserts lock both positive (new pattern present) AND negative (pipe-to-head absent) in both .sh and .ps1

Cross-OS validation

• Linux (this PR author's primary): empirically validated, SIGPIPE eliminated.
• Windows: PS1 change is comment + bash one-liner rewrite to mirror Linux form. Symptom never observed there (no pipefail in bash -c sub-context), so this is parity-only. Empirical Windows validation tracked in WIN-002-windows-smoke-sweep spec (full clean-VM sweep).