docs(backlog): the raised rail earned itself in eight minutes

[G1NG4R]

Backlog only — no code. The last swap of the night, and the measurement that
justifies the change made during it.

What it carries

• The raised witness rail earned itself, by 8.5 minutes. e2e on main ran
  38:30 and green — a new maximum. The witness started at 01:19:26 with
  WITNESS_MAX_SEC=2700; the old 1800 s rail would have expired at 01:49:26 and
  the deploy came at 01:57:53. That measurement would have been lost the way
  chore: bump traefik from 24841fe to 575fa15 #439's was, for a different reason: there the runner queue sat in front of the
  job, here e2e itself is longer than the rail.
• The lesson is distrust of the band, not the band. A sampled upper bound was
  overtaken twice in one night: 17:38, then 34:57, now 38:30. Start as soon as
  e2e is in_progress and set the rail generously — starting early costs
  nothing, a short window costs the whole measurement.
• The swap: a60f44c, ops.lastDeploy.at 01:58:13Z, serving startedAt
  01:58:28.779Z, offset +15.8 s, 2359 × 200 per path, zero non-200.
  The identity was read ten times and was identical ten times — after the
  twin-read finding in the entry above, it is not read once any more.
• test: the witness tail is measured from the twin, and closes before #304's window #453 gains a fifth data point: twin gap 13.8 s, coverage 16.2 s, so this
  window does cover the known bug: a container swap sometimes drops one request, and two samples cannot say how often #304 offsets. Four of five do; 3b990e5 with
  3.4 s remains the outlier.
• I1 · Boot-Sequenz is written up as the next unit, with the three decisions
  it waits on: new --d-boot-* tokens (the offsets cannot be hard-coded as
  animation-delay — tools/check-tokens.sh:80 catches that), Lenis and GSAP
  for I2 against chore: initial JS is 143 KB of a 150 KB budget, before a page exists — due with L8 #237's bundle budget, and the ambient layer sitting in the
  sheet's spec table while the build plan defers it to I3.

Tally: 6 drops in roughly 37 800 requests, twelve swap runs in a row without
one. #304 gains no new evidence — five witnessed swaps tonight, all clean.

make check green.