docs(backlog): the tail is thirty seconds minus the twin

[G1NG4R]

Backlog only — no code. The two measurements the 50bd9cc merge produced,
written while they were fresh.

What it carries

• The witness tail is 30 s minus the gap between the twin and the final
  process. Measured twice in one evening: a 26.6 s gap left 3.4 s of
  coverage after the serving process came up, a 15.6 s gap left 14.4 s. The
  first misses both known bug: a container swap sometimes drops one request, and two samples cannot say how often #304 offsets (+12 s, +12.9 s), the second clears them
  barely. Coverage is a remainder of how long the rollout takes, not a property
  of the instrument — which is what test: the witness tail is measured from the twin, and closes before #304's window #453 is for.
• e2e went red on main at the chore: two e2e tests flicker, and the deploy gate hangs on the suite they are in #429 count test.
  touch-targets.coarse.spec.ts:211 at coarse-390, 2 interactive elements
  expected in <main> on / and 1 counted; 1 failed, 348 skipped, 2457 passed
  — the same line and the same numbers as docs(backlog): the red second is the idle drop, not the swap #448. One gh run rerun --failed,
  green, no code touched. The e2e band held to the second: 13:35 and 17:38.
• The swap itself was clean. ops.lastDeploy.at 23:32:18Z, startedAt
  23:32:29.559Z, offset 11.6 s, and the page reports v0.47.2-12-g50bd9cc at
  p95 73.3 ms.
• pkill -f matches its own command text, exactly like pgrep -f. The
  guard loop was killed with a pattern that appears in the same line, so it shot
  its own shell before the kill <pid> beside it ran. Use the PID you already
  hold.

The idle-drop tally gains two rows: the 1800-sample window with no deploy in it
(deploy was gated off by the red e2e, so its ✗ verdict is correct), and
140 × 200 per path, 0 drops over the swap.

make check green.