The roll rehearsal's real two-process test is green under the gate's load, or quarantined under its own issue #2032

I widened the timing margins in test/rehearsal.test.ts instead of quarantining the file. bun scripts/flake-hunt.ts admit test/rehearsal.test.ts now shows ten green runs in random order under load. The changes are in the test only:

• Linger: the lingering test now runs the old forge with a 5 s linger, up from 2 s. A busy machine would have to stall the probe loop for over 4.7 s to put "left rotation" after "stopped answering". Before, a 1.7 s stall was enough. The exit-time assertion moves from at least 2000 ms to at least 5000 ms to match.
• Window: the rehearsal now runs 30 s, up from 20 s. A slow boot of the new forge no longer eats the window the load needs for the roll.
• Assertions: the exits-at-once test is unchanged and still expects refusals. The lingering test still expects zero refusals, traffic to both forges, and the old forge leaving rotation before it stops answering.

I didn't reproduce the original failure. Three plain runs on an idle machine passed before I changed anything. The cost is a slower file, about 60 s per run.

Admission output:

admission: test/rehearsal.test.ts — 10/10 green, random order, under load
  run 1: seed 1958367548 green
  run 2: seed 1104356147 green
  run 3: seed 1636047315 green
  run 4: seed 2044224590 green
  run 5: seed 2998538531 green
  run 6: seed 507488893 green
  run 7: seed 3678750070 green
  run 8: seed 3121875993 green
  run 9: seed 1830613977 green
  run 10: seed 3558062952 green

test/check-list.test.ts passes (8 tests). To try it, run bun test test/rehearsal.test.ts.

Closes #2020