Cassette Build Report 021 — The Review Gap Became a Work Order

Scope note: This report covers the S06 handoff from a Linux review to an eligible Mac execution, including the failed attempts that produced no evidence. It does not treat a model’s confidence or a copied test packet as a substitute for running the named checks.

There is a particular kind of failed review that sounds productive. The agent explains the plan, names the right files, and never reaches the machine that could answer the question.

S06 gave me several of those. Opus 5 Extra could inspect the repository from Linux, but it could not open a terminal on the Mac or run the APFS and F_FULLFSYNC checks. The first sub-agent returned no action ledger. Kimi K3 announced a careful plan and called a nonexistent Bash() function three times. OpenClaw with GPT-5.6 Terra reviewed a packet I had assembled and then returned four passes without being allowed to inspect the checkout. None of those runs changed the evidence.

I finally said, “I am just looking for one solution, not two more models.” The useful answer was not another model. It was a finite handoff that named the nine claims Linux had not reached and gave them to a reviewer with the right machine.

The handoff listed the expected commit, branch, dirty files, temporary directory, cleanup rules, and the exact APFS boundaries to execute. It told the Mac reviewer what it could not claim: no physical sleep, no USB bus reset, no cable movement. The earlier review remained authoritative for the portable logic. The new run existed only to fill the environmental gap.

The fixture makes one of those boundaries concrete:

observed = process.stdout.readline().strip()
if observed != f"BOUNDARY:{boundary}":
    raise AssertionError(f"expected {boundary!r}, received {observed!r}")
process.kill()
assert process.wait(timeout=10) == -signal.SIGKILL

Those lines are in tests/test_s06_transactions.py. A label in a report is not proof. The child process has to reach the label, die, and be observed as dead before the parent reader is checked.

Claude Code ran the handoff on the Mac. It observed twenty-eight writers reaching their named boundaries and exiting by SIGKILL, thirty-nine APFS attaches and detaches, and 5,600 reads from twenty-eight concurrent parent readers. It traced the hooks into production code, confirmed that the fixture had not skipped its work, and verified F_FULLFSYNC on the mounted image. The opening and closing Git states matched. Its temporary sampler was removed. No disk image remained mounted.

The report also corrected three small claims in the closeout: one transition advanced the journal but was not a production action, one rollback generation lacked an independent recomputation, and the garbage-collection assertion checked presence rather than the exact removed set. That mattered. The eligible reviewer did not simply repeat the verdict I wanted.

The gap prompt did not close S06 by itself. The Mac execution did. The prompt made that execution possible without making me translate “Linux cannot do this” into another round of vague delegation. A good handoff is not a longer review. It is the shortest list of facts another environment can actually prove.