fix(codex-implementer): stop lanes colliding in /tmp and stalling forever

[wooter]

Three defects in the lane, all observed repeatedly across parallel lanes on macOS. Independent of #12 — different hunks, the two merge cleanly in either order.

1. mktemp -t codex-spec.XXXXXX doesn't do what it looks like on macOS

BSD mktemp treats -t's argument as a prefix, not a template. The literal XXXXXX survives into the filename:

$ mktemp -t codex-spec.XXXXXX
/var/folders/ph/1j0k93y96x19llq06r2m0pn40000gn/T/codex-spec.XXXXXX.TJEjQBKZt2

So every lane on the host emits a path matching the same codex-spec.XXXXXX.* glob. That is the enabling condition for cross-lane pickup: a lane recovered its spec path by globbing and ran a different, concurrently executing lane's spec. I found two such stray files sitting in TMPDIR on my machine from real runs.

The block sits directly beneath the words "never a fixed path (parallel lanes on fixed paths corrupt each other)" — the intent is already right, the code just doesn't deliver it. Passing an explicit template as an operand substitutes the Xs on BSD and GNU alike.

Fix: one private scratch dir per lane, with spec.md and final.txt inside it.

2. Nothing said how to carry the path across tool calls

$SPEC only lives for one Bash call, so lanes improvise:

• ls /tmp/codex-spec.* | tail -1 — sorts by random suffix, not mtime, so under concurrency it cheerfully returns another lane's spec
• /tmp/.codex_spec_path — a fixed global sidecar; two lanes clobber it deterministically, not occasionally

Both are now called out explicitly, with the two safe options (same tool call, or echo the path and copy it forward).

3. Lanes background codex and then wait forever

The dominant failure in my logs — eight separate incidents. A lane launches codex exec as a background task, ends its turn "waiting for the completion notification", and nothing is wired to deliver that notification to a subagent. Observed shape: 60–100k tokens across dozens of tool calls, empty working tree, no report, while the caller believes work is in flight. One session hit it three times.

Related: a 600 s shell wrapper is meaningless when the Bash tool call itself defaults to 120 s. The tool kills the command long before the documented cap can apply, so the "ten-minute wall clock" is fiction unless the tool-call timeout is set to match.

Both are now stated plainly in step 2, plus a hard rule: never end your turn with a codex process still running.

Verification

Check	Result
mktemp -d "${TMPDIR:-/tmp}/codex-lane.XXXXXX" on BSD mktemp	Xs substituted, no literal XXXXXX
two simulated concurrent lanes	distinct dirs, neither spec swapped
overlap with #12	none — disjoint hunks

Docs-only change.