hertz -> doyle + todlando + liam. STOPPED-CELL PREDICATE REPORTED AS WRITTEN, AND IT IS WORSE THAN A
DURATION ASSERTION. Fixed before the arm, not during it. New blob named below.

1. THE PREDICATE AS WRITTEN, verbatim from twohost_bootstrap.rs @1b636dc2, the state it would have run in:
     fn bootstrap_stopped_from_second_machine() {
         let budget = Duration::from_secs(1);
         let (face, elapsed) = rig.attempt("stopped/attempt1", budget);
         match face {
             Face::Stopped => eprintln!("refused in {} ms - STOPPED is distinct from ADMISSION-BLOCKED"),
             Face::Blocked => panic!("expected STOPPED, saw ADMISSION-BLOCKED (timed out at {} ms) -
                 THIS IS THE FINDING, not a flake: the two refusals share a face at the wire ...
                 Report the measured times; do not soften this assertion."),
   There is NO assertion on elapsed anywhere in the cell - todlando was looking for the right thing and it
   is not there. The sub-second bound is the CONNECT BUDGET ITSELF, which is strictly worse: a refusal
   arriving at 2 s does not merely red, it is classified Face::Blocked and reds AS THE LEG'S HEADLINE
   FINDING, under a message that instructs the reader not to soften it. That is the sidecar-parser class
   escalated one level - a predicate that cannot express the right answer, wired to the conclusion we
   most wanted to be able to trust.

2. WHERE THE ~2 s LIVES - one candidate eliminated, by measurement, on my own box only. todlando's two
   numbers (2.042 s loopback, 2.023 s own-LAN) are both Windows-CLIENT measurements, so they cannot
   separate "Windows is slow to emit RST" from "curl/Windows is slow to report one". They also do not
   transfer to my cell, whose client is Linux and whose library is not curl. So I measured MY client's
   contribution locally on kitsubito - a connect to a closed localhost port, no LAN, no Windows, an
   unmodified subject that references nothing in the rig:
     local_refused attempt0  0.242 ms   attempt1  0.031 ms   attempt2  0.027 ms  (ConnectionRefusedError)
   My client contributes ~0. So the ~2 s is on the peer side of the wire in some form, and which form is
   still open. I am NOT proposing an arm for it - see 4, it answers itself for free.

3. FIX APPLIED, smallest diff that is correct under BOTH remaining readings: budget 1 s -> 6 s, matching
   the blocked arm. A genuine refusal returns the instant the RST lands whatever its latency, so raising
   the budget cannot cost discrimination; only a true drop now consumes the full 6 s. The discriminator
   stays the FACE (RST vs timeout) and elapsed stays logged-never-asserted. The comment records both
   measurements by hand and date so the next reader cannot re-tighten it.
     blob 1b636dc235908f3b6217feb780c1d9af081c9bbd -> 49102e395c3778a8378f3cc5ed06b95ac5a00bbe
   check/clippy --workspace --all-targets running now; I will name the exit codes before I fire, and the
   population must still read 4 cells. If either is nonzero I do not fire.
   NOTE so nobody misreads it later: 6 s is now BOTH arms' per-attempt budget, so a stopped-timeout and a
   blocked-timeout take the same wall. That is intended - the wall was never the discriminator - but no
   one should read the 6 s as distinguishing anything.

4. FREE DATUM FROM A3b, not a criterion. If my cell passes it logs elapsed_ms beside the face. A few ms
   there proves the ~2 s is the Windows CLIENT path and Windows emits RST promptly; a ~2 s reading proves
   the peer is genuinely slow to refuse. Either way the cell PASSES and liam's residual question about
   where that 2 s lives gets answered by an arm already in the sequence, at no extra cost.

5. Labelling, per doyle's (4): my stopped cell runs at F-PRESENT state and is labelled THROUGH THE HAND
   RULE exactly as the admitted cell was. It witnesses the product's refusal semantics; it does not
   witness the product's admission, and no green from it may be cited for REQ-WEB-LAN-BOOTSTRAP-FIREWALL's
   admitted leg. Same sentence, same place, in the source paragraph todlando asked me to keep.

Standing: fire only on todlando's stamp, after liam's F re-add. Nothing of mine on HFENDULEAM.