test(autonomy): five low-severity items from the lane-stop-gate arm-claim verification

[kyle-sexton]

Five low-severity items surfaced by the fresh-context verifier on #2065 (lane-stop-gate arm-claim
race). All outlived that merge; none is blocking. Filed together because they share one file and one
review pass.

The verifier's overall verdict on #2065 was DISCHARGED, evidenced by a clean 83/0 run on the branch,
82/1 with only the type-guard line deleted, and all five rc0 sites red when the hook is made to exit
127. These are what it found alongside that.

Test-harness defects — plugins/autonomy/hooks/lane-stop-gate.test.sh

1. :804 — unchecked mkfifo can produce a false PASS. Demonstrated by stubbing mkfifo to
   fail: the case still reported PASS. A test whose setup can fail silently is not a test.

2. :820 — kill -9 on the failure path orphans the grandchild. The FIFO case deliberately
   hangs the hook (that is the bug it proves); killing the subshell does not reap the grandchild
   blocked in open(). Observed live during verification — idle processes blocked on a FIFO in an
   already-deleted temp dir, which then survive the run.

3. :699 — rc0 silently passes on an empty argument. Latent: no live call site passes one
   today, so this is a trap for the next test author rather than a current false PASS.

Runtime observations

4. umask 077 is a no-op on MSYS, so the claim file lands 644. The claim sidecar's confidential-
   ity does not depend on mode — its content is a session id, and the exclusive-create is what
   provides the guarantee — but the umask line implies a protection it does not deliver on Windows.
   Either make it real or say it is advisory.

5. A mid-lane downgrade to 0.13.0 can hand a bound record to a second session. The compatibility
   read added in 0.13.1 makes an older record's in-record owner authoritative, but a lane that
   downgrades mid-flight reaches code with no sidecar concept while a sidecar exists. Narrow window,
   operator-initiated, and the fail direction is over-gating rather than under-gating.

Suggested handling

1-3 are test-quality fixes and can land together in one pass. 4 is a one-line honesty fix on a
comment or a real chmod. 5 is worth a documented note rather than code, unless mid-lane downgrade
is a supported operation.

Provenance

Found during the stranded-review-findings sweep verifying PRRT_kwDOTCGFQM6WACHN (#1865) on PR
#2065. Deliberately not folded into that PR: none discharges the finding under repair, and the
branch was already carrying a P1 race fix that deserved a clean diff.

[kyle-sexton]

Claimed by cloud agent; Fixes PR incoming from branch cursor/*-63e1.