docs: wire detached-eval guidance into /ship + correct CLAUDE.md flags

- /ship eval step (sections/tests.md): long eval suites launch via gstack-detach
  (own session, machine lock, EXIT sentinel) so a turn boundary can't kill a
  30+ min run mid-ship — the exact failure observed during this branch's ship.
- CLAUDE.md: correct the now-stale /tmp reference; document the --lock (serialize
  worktrees, no API saturation), --timeout watchdog, run-scoped log, and the
  guaranteed EXIT sentinel the poller breaks on.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

CLAUDE.md

@@ -839,13 +839,20 @@ the run can also die to idle-sleep. `gstack-detach` fixes both: a fresh session
 (escapes the group SIGTERM) wrapped in `caffeinate -i` (blocks idle-sleep).
 
 - Use the `eval:bg*` scripts (`eval:bg`, `eval:bg:all`, `eval:bg:gate`,
-  `eval:bg:periodic`) — they wrap the eval command in `gstack-detach` and stream to
+  `eval:bg:periodic`) — they wrap the eval command in `gstack-detach` with the
-  `/tmp/gstack-evals.log`. Or call `gstack-detach <log> -- <cmd>` directly for any
+  machine-wide `gstack-evals` lock (concurrent worktrees serialize instead of
-  long agent job. Export `ANTHROPIC_API_KEY` first (never pass keys in argv).
+  saturating the shared model API), a per-tier watchdog, and a **run-scoped** log
-- Then **poll the logfile** with a death-aware watcher: break on a `### DONE ###` /
+  under `~/.gstack-dev/eval-runs/` (no shared-`/tmp` collision). Each prints its
-  `EXIT=` marker OR on "runner process gone before DONE" (silence is not success —
+  log path. Or call `gstack-detach [--lock NAME] [--timeout SECS] [--label LBL] --
-  cover the crash/kill path, not just the happy path). The detached run survives
+  <cmd>` directly for any long agent job. Export `ANTHROPIC_API_KEY` first (never
+  pass keys in argv).
+- Then **poll the printed logfile** with a death-aware watcher: break on the
+  guaranteed `### gstack-detach EXIT=<code> ###` sentinel (success AND failure are
+  both marked, so silence is never mistaken for success). The detached run survives
   even if your watcher gets reaped, so re-checking the log always works.
+- Why the lock: a shared dev box with several Conductor worktrees will rate-limit
+  the model API if two eval suites run at once (15-way concurrency each), which
+  mass-times-out E2E tests. The lock makes the second run WAIT, not collide.
 - Humans running `bun run test:evals` foreground in their own terminal don't need
   this — Ctrl-C is intended there. Detachment is for agent-launched runs only.
 

ship/sections/tests.md

@@ -332,6 +332,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`~/.claude/skills/gstack/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+~/.claude/skills/gstack/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.

ship/sections/tests.md.tmpl

@@ -76,6 +76,22 @@ EVAL_JUDGE_TIER=full EVAL_VERBOSE=1 bin/test-lane --eval test/evals/<suite>_eval
 
 If multiple suites need to run, run them sequentially (each needs a test lane). If the first suite fails, stop immediately — don't burn API cost on remaining suites.
 
+**Long eval suites (30+ min): launch detached so a turn boundary can't kill them.**
+A plain backgrounded eval lives in the harness's process group and dies to a
+SIGTERM ("polite quit") on a turn boundary, a stopped monitor, or an interruption
+(observed mid-`/ship`: `script terminated by signal SIGTERM`). Run it through
+`~/.claude/skills/gstack/bin/gstack-detach` instead — it survives in its own
+session, serializes against other worktrees via a machine lock (no API
+saturation), and writes a guaranteed `### gstack-detach EXIT=<code> ###` sentinel:
+
+```bash
+~/.claude/skills/gstack/bin/gstack-detach --label ship-evals --lock gstack-evals --timeout 5400 -- <project eval command>
+```
+
+Then poll the printed log path; break on the `EXIT=` sentinel (covers both pass
+and crash — silence is never success). The detached run survives even if your
+poller is reaped.
+
 **4. Check results:**
 
 - **If any eval fails:** Show the failures, the cost dashboard, and **STOP**. Do not proceed.