…red rulebook, the key vocabulary, the self-registering sweep

The fence that retires the legacy inventory.lock is a read-only file
swapped in by rename. A process that checked the fence just before the
swap then opens the fenced inode for writing and gets EACCES, which the
retry path did not catch: it escaped as a raw PermissionError. CI hit
the interleaving once under coverage instrumentation. The open now
re-verifies the fence and retries through the new lock generation; a
genuine permission problem still surfaces. A deterministic test pins
both sides.

The client treats the id reply as receipt and the test asserts the
moment the push returns, so a rename appended after sendall loses that
race on a loaded machine. One matrix job caught the interleaving; the
handler now appends before it replies, which makes the receipt a
guarantee instead of a hint.

Every job now carries a timeout near double its slowest healthy run, so
a wedged runner dies in minutes instead of holding the run for the
six-hour default and mailing a failure nobody caused. The upstream
shpool test suite gets one retry: its cleanup_socket test is timing
flaky on loaded machines, and one stumble of someone else's test should
not paint our run red. Our own tests still fail on first strike.

Version references and the changelog heading name the release being cut:
the picker no longer loses a keypress under load, one shared rulebook
reaches every profile, the key vocabulary says quit where it means quit,
the sandbox sweep registers its own prefixes, and losing the
publishing-lock fence race retries instead of crashing.

A visitor used to read four paragraphs of identity model, colour
allocation and subscription behaviour before reaching the screenshot,
which made a small tool read as a systems project. The picture now sits
under a two-line explanation, the reason the thing exists comes next in
the author's own words, and the implementation detail moved into the
sections that already covered it. The beta block became one line at the
top with its promises kept in full further down.

Two corrections rather than styling: the suite claim said 2,700 tests
run on both platforms for every change, and macOS runs a selected set of
modules, so it now says native Linux and macOS coverage in CI. The page
also no longer surveys other session managers or tells a reader another
tool would serve them better.

…terval

The argv helper waited five seconds for a separate watcher process to
write its argv file, then returned an empty string. On a loaded runner
the watcher had not started in that window, so the caller's assertion
reported that '--interval .25' was missing from '' and pointed at the
interval value rather than at the process that never ran. CI caught it
in the coverage job. The wait is now thirty seconds, a half-written file
does not count as an answer, and running out of time raises with the
reason.

…get back

The tagline names the trade the tool actually makes: a terminal window
is a view, a session is the work, and closing the first does not touch
the second. The line under it says what you get instead, in the three
words that matter, and the third says why closing is safe. The identity
model, the colour rules and the subscription system stay where they
were, further down.