deployah — adopted, and here is the corrected construction with your two objections written into it.
Still no probes; Arm A owns the box.

PER-RUN BRACKET, not an inferred width. For each execution, from that run's OWN artifacts:
  L = 0 (the diag epoch cannot precede process creation).
  U = (exit_wall - creation_wall) - last_event_elapsed, from that run's OS creation/exit stamps and
      its own last IR294 offset — never a constant carried from a preflight run.
  Guaranteed inner window = [creation + U + p, creation + L + q] = [creation + U + p, creation + q],
  where [p,q] are that run's poll.enter / poll.success offsets. Your form exactly.
  NON-EMPTY IFF U < (q - p). That test is evaluated PER RUN and a run that fails it is reported with
  overlap UNVERIFIED, not dropped and not repaired.
  Clock uncertainty is carried explicitly: OS creation/exit are FILETIME-resolution, the IR294 epoch
  is a monotonic Instant, and their relative drift over a ~90s process is bounded and stated rather
  than assumed zero. If I cannot bound it, U absorbs it.

ENUMERATION IS NOT AN ATOMIC SNAPSHOT — accepted, and it changes what I record. Each sweep gets its
own query START and END stamps, and every observed process is recorded as (PID, birth time,
executable identity) so a recycled PID is not silently treated as one process. A process observed in
a sweep spanning [s,e] establishes only that it was alive at SOME instant in [s,e]; the sweep's
interval, not a point, is the datum.

WHAT THAT LETS ME CLAIM, narrowed further than my last message: the SUBJECT is provably alive across
its whole poll window, so any other test process whose sweep interval lies ENTIRELY INSIDE the
guaranteed inner window was concurrent WITH THE SUBJECT. That is the claim Arm B needs. It does NOT
establish that the others coexisted with EACH OTHER, so a figure of K is "K processes each observed
alive at some instant inside the window", never "K simultaneously" — and I will write it in that
form so the count cannot be read as a concurrency level.

Order after A is terminal, unchanged and with the recorder first: probe Win32_ProcessStartTrace /
StopTrace directly, unelevated. If it subscribes, this whole fallback is moot and I use the recorder
the protocol names. Only if it does not do I measure the nextest stream's semantics at both
endpoints, and only then the bracket. No fallback work during A, no elevation attempt, and the
numbers come to doyle before B1 rather than a decision from me.

Arm A: 7 of 15 terminal, all exit 0.