workflow-fix: document the 'CMD -c ... || echo 0' double-print trap in gotchas.md

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a near-miss on task #1773 (2026-07-31). A pod-release watcher silently never fired because of a shell idiom that double-prints. The pod would have billed 11 extra hours of a 1x H100 ($25-45) before the watcher's 20h timer exited on its deliberate never-terminate-on-a-timer arm. It was caught only because a live agent probed the pod by hand.

Goal

Add a .claude/rules/gotchas.md entry for the CMD -c ... || echo 0 double-print trap, so the next liveness/termination guard does not silently no-op.

Workflow gap

• Bug observed: issue1773_fulldict_watch.sh:52 computed launcher liveness as ALIVE=$(podssh "pgrep -c -f 'PAT' 2>/dev/null || echo 0" || echo "unknown"). pgrep -c PRINTS 0 and also exits 1 when nothing matches, so the || echo 0 fallback fires as well and $ALIVE becomes the two-line string "0\n0". The gate [ "$ALIVE" = "0" ] therefore never matches, and the watcher concludes the launcher is still alive forever. Confirmed on #1773 via cat -A (the log line is split across two physical lines) and a direct pod probe showing pgrep -c printing 0 with rc=1.
• Why it is a workflow gap: the trap is invisible on inspection — the code reads as a standard defensive fallback, and it fails OPEN in the most expensive possible direction (a resource-release guard that never releases). .claude/rules/gotchas.md is the documented home for exactly this class of trap and has no entry for it. The idiom is not #1773-specific: it bites whenever a command BOTH prints a count AND exits non-zero on empty, which is pgrep -c, grep -c, and anything wrapping them.
• Confidence (emitter): high — root cause identified, reproduced by direct probe, and the money consequence is quantified.
• verified-at-filing: grep -rn "pgrep -c" scripts/*.sh scripts/*.py .claude/rules/*.md .claude/agents/*.md -> 0 hits (2026-07-31), and grep -ci "pgrep -c" .claude/rules/gotchas.md -> 0. So the trap is currently UNDOCUMENTED and no committed workflow-surface file uses the idiom; this filing is preventive, not a live-red fix. The offending script is a per-issue driver (scripts/issue1773_fulldict_watch.sh), which is deliberately OUT of the workflow-fix surface and is not proposed for change here.

Proposed change (candidate diff sketch — refine in planning)

Add to .claude/rules/gotchas.md, in the shell/ops section:

+**`CMD -c ... || echo 0` double-prints.** `pgrep -c` (and `grep -c`) PRINT `0` AND
+exit non-zero when nothing matches, so a `|| echo 0` fallback fires too and the
+captured value becomes the two-line string `"0\n0"`. Any `[ "$X" = "0" ]` gate then
+never matches. This fails OPEN, so it is worst in liveness / termination guards:
+#1773's pod-release watcher (2026-07-31) never fired and would have billed ~11 extra
+hours of a 1x H100. Use `X=$(CMD -c ... 2>/dev/null); X=${X:-0}` or `X=$(CMD ... | wc -l)`,
+and diagnose a suspected case with `cat -A` (the value shows as two physical lines).

Scope / surfaces

• Primary target: .claude/rules/gotchas.md
• Consider whether the LESSONS.md index row for gotchas needs its trigger widened to name resource-release guards; the planner decides.

Constraints / invariants

• Documentation-only change to a rule file. No behavioural code change is proposed: the offending script is a per-issue driver, out of the workflow-fix surface by design.
• scripts/workflow_lint.py no-flags run passes; --check-lessons-index stays green.
• This session runs under the recursion guard and did not auto-route its own candidates.

Provenance

• workflow_fix_target: .claude/rules/gotchas.md
• fingerprint: pgrep-c-double-print-release-guard

Origin: near-miss on #1773's full-dictionary run, 2026-07-31. Diagnosed by the run's implementer, which correctly declined to file (the offending file is a per-issue driver and therefore out of scope); the orchestrator files this because the generalizable trap belongs in the always-available gotchas rule, which IS workflow surface.