The `clide` shell client only ever computed the socket path from the
workspace's FNV-1a hash — it never read CLIDE_SOCK, so the var the app
exports to spawned agents was a no-op and a bogus CLIDE_SOCK was silently
ignored (it still hit the discovered instance). Now: when CLIDE_SOCK is set
the client connects to it directly, beating discovery, and a dead socket
aborts with EX_UNAVAILABLE rather than falling back to a different instance
(the split-brain footgun). Unset → the deterministic per-workspace path
(D-70) as before. The server keeps deterministic binding; this is a
client-side targeting override only.
Also fixes a pre-existing -Wstringop-truncation warning in
find_workspace_root (strncpy+manual-null → snprintf).
Tests: the e2e suite now clears the inherited CLIDE_SOCK so discovery tests
stay hermetic (the suite may run inside a clide instance), plus two new
tests — a valid CLIDE_SOCK pins from a non-git dir, a dead one fails loudly
and returns no data.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jpmschweitzerandClaude Opus 4.8
6af980f31c