fix(install): stop legacy commands cleanup from deleting itself on case-insensitive filesystems

[edheltzel]

Problem

Every fresh ./install.sh run left every Recall:* slash command missing, and re-running the installer reproduced the exact same failure forever — only recall doctor --fix actually worked, until the next install.sh run undid it again.

Screenshot reproduced:

✗ Install verification failed — these symlinks are not in their expected state:
    MISSING:   /Users/ed/.claude/commands/Recall/add.md
    MISSING:   /Users/ed/.claude/commands/Recall/doctor.md
    ... (all 9 commands)
✗ Self-check failed — the install did not fully verify; left marked incomplete.

Root cause

install.sh's Commands step creates the commands/Recall/*.md symlinks, then a "legacy lowercase directory" cleanup runs right after to remove any leftover commands/recall/ from an older naming scheme:

local commands_legacy="$CLAUDE_DIR/commands/recall"
...
if [[ -d "$commands_legacy" && "$commands_legacy" != "$commands_dest" ]]; then
  rm -rf "$commands_legacy"
fi

On the default case-insensitive-but-case-preserving macOS filesystem, commands/recall and commands/Recall are the same directory on disk. The plain string != sees them as different paths, so the guard never trips — rm -rf deletes the directory the loop directly above had just populated with symlinks.

The identical block was duplicated in lib/install-lib.sh's recall_copy_runtime_files (used by update.sh), so update.sh had the same bug.

recall doctor --fix "worked" only because its repair path has no such cleanup step — but the next install.sh/update.sh run re-triggered the delete.

Fix

Replace the string comparison with bash's -ef operator (compares by device+inode, so it correctly identifies "same directory on disk" regardless of case-folding) in both duplicated locations.

Testing

• Added a regression test in tests/install/update.test.ts exercising recall_copy_runtime_files against a case-aliased legacy path.
• Reproduced the original bug against pre-fix code (symlinks vanished after the Commands step) and confirmed the fix prevents it, via direct function-level testing against lib/install-lib.sh.
• Full real end-to-end install.sh --yes dry run (isolated sandbox $HOME): self-check passes, all 9 command symlinks survive two consecutive runs (matching the original bug's exact repro).
• Ran the real, fixed install.sh against my own live $HOME: cleared a stale .install-incomplete marker left over from hitting this exact bug earlier, recall doctor health score went from 26/30 to 27/30 with "Install completed (no interrupted run)" now passing.
• Full test suite: 1198 pass; 10 pre-existing failures unrelated to this change (OpenCode/Pi MCP JSON-config hardening tests).

🤖 Generated with Claude Code