docs(backlog): roughly one swap in six, and the clean one counts

[G1NG4R]

Backlog only — no code. The last measurement of the run, and it is left for
you to merge rather than merged: merging it produces the next swap, which would
need witnessing in turn, and the loop ends here.

What it carries

• The swap right after the 503 was clean — 72f6a8e, 1103 × 200 per path —
  and it counts, because the window reached 15.2 s past the moment the
  serving process came up. A hole in the measured +11.3…13.3 s band would have
  been inside it. It was not there.
• A bounded rate for bug: a container swap sometimes drops one request, and two samples cannot say how often #304, which its title says it lacks. Across seven
  witnessed swaps, six had coverage wide enough to see a ~12 s hole, and one of
  those six showed it: roughly one swap in six. The table is in the entry and
  as a comment on bug: a container swap sometimes drops one request, and two samples cannot say how often #304 with the same arithmetic.
• Two caveats kept in the text, because both weaken the denominator and
  neither weakens the numerator: 944a1e4's 13.3 s coverage sits exactly on the
  band's upper edge, and d0bdd76's coverage is derived rather than read — the
  runner's printf began with a dash and never wrote its exit timestamp.

make check green.