docs(backlog): the hole sits twelve seconds after the process

[G1NG4R]

Backlog only — no code. The first non-200 of the night, and what it cost to see
it.

What it carries

• A fourth observation for bug: a container swap sometimes drops one request, and two samples cannot say how often #304. The 206df65 swap served one 503 on /
  while /api/health held 200 through the same second: 1107 requests per path,
  1106 × 200 and 1 × 503. The first non-200 in roughly 42 500 requests measured
  overnight across six witnessed swaps.
• Converted to wall clock, as the rule demands. The sample label is
  wall-clock, not an index (elapsed=$(( now - start + 1 )),
  tools/witness.sh:316), so the red second is pinned to 05:14:37Z ±1 s. The
  serving process came up at 05:14:25.711Z, which puts the hole at +11.3 to
  +13.3 s after it — and the three solid offsets are now a cluster around
  12 s (+11.3…13.3, +12, +12.9) rather than three scattered points. Same shape
  as before: a 503 rather than a 404, and only the container serving the page.
• It was visible only because the remainder was large enough. This window
  covered 15.3 s past the serving process; the 3b990e5 swap earlier the
  same night covered 3.4 s, because the 30 s tail runs from the twin. A
  hole at +12 s is invisible to a 3.4 s remainder, and one of the six windows
  would have swallowed this 503. test: the witness tail is measured from the twin, and closes before #304's window #453 is the difference between measuring this
  defect and not — argued from last night's own two outcomes.
• A correction to the tally, against my own earlier wording. The escalation
  criterion is "two failures on one path or one that falls in the swap".
  This one falls in the swap, so it is met for the first time since H7a — all
  six previous failures sat at rest, seven to eighteen minutes away from a swap.
  The issue it calls for already exists (bug: a container swap sometimes drops one request, and two samples cannot say how often #304); the sentence below the table no
  longer reads "not met", and the docs(backlog): #431 accepted, and the thirteenth swap has no hole #444 case is kept separate so the two do not
  contradict.

The observation itself is on #304 with its arithmetic beside it, not only here.

make check green.