Phase 12.L.E.7.5 — fix: version subcommand fast-path stops marker deletion
Summary
Root cause finally pinned via J.L.E.7.4 diagnostic dump.
The .sh's Phase B post-restart sanity check runs:
timeout 5 "$INSTALL_DIR/squid-tentacle" versionsquid-tentacle is a symlink → Squid.Tentacle (our placeholder bash). Without version argument handling, the script falls into the sleep loop. After 5s timeout, SIGTERM fires the cleanup() trap which removes the marker file — even though the systemd service itself wrote it correctly seconds earlier.
Trace (from diagnostic dump)
| Step | State |
|---|---|
| Phase B mv swap | InstallDir has v2 contents (version.txt = 2.0.0 ✓) |
| systemctl restart starts v2 | v2 service writes marker = "2.0.0" |
| healthz curl passes | python3 responder works |
| version probe runs symlink chain | timeout 5 → SIGTERM → cleanup → rm marker |
| .sh writes SUCCESS, exits 0 | |
| Test reads marker | ABSENT |
Fix
Handle version subcommand fast-path in the test service script. When $1=version, print version.txt content + exit 0 immediately. No sleep loop, no trap fire, no marker removal.
This mirrors what the real Squid.Tentacle CLI does (per the .sh's comment: "use the version subcommand (added in the same Phase 2 Part 2 follow-up commit that added VersionCommand.cs)").
Lesson
No mock or unit test would have surfaced this. The bug is the interaction of:
- Production .sh's exact version-probe argv invocation
- Bash placeholder running as the "agent binary"
- Bash trap on SIGTERM cleanup
- Marker removal happening AFTER service-side write
Only real systemd + real chain + real timeout 5 could expose it. J.E.3.1 pattern continues — high-fidelity E2E earns its keep one bug at a time.