test(terminals): count the shared kill_all grace in clock reads, not seconds

[nt-nerdtechnic]

Summary

• test_kill_all_waits_the_longest_grace_once_not_once_per_pane failed on Windows backend 4/4 (run 38443251695, job 115454959944): three panes took 10.66s. The log shows all three sessions closed with exit=1 — on Windows TERM terminates the child, so no grace ran at all; the 10.66s was the ps snapshot, breakaway sweep and lifecycle drain on a loaded runner. Test misjudgment, not a product defect.
• Reuses _SteppingClock from test(terminals): count the spec-less kill_all grace in clock reads, not seconds #204: a kill_group spy records the clock at each TERM and escalation. Asserts every TERM precedes the first escalation and each escalation lands one grace (within 3 clock steps) after the last TERM. On Windows (no escalation) it asserts only that the platform is win32, as test(terminals): count the spec-less kill_all grace in clock reads, not seconds #204 does.

Verification

• Reproduced locally by making _ps_snapshot sleep 7s: the old test fails with the same message (10.06s); the new test passes.
• Mutation: kill_all waiting one grace per pane sequentially → escalations at 3.1 / 6.2 / 9.25 and the new test fails.
• Whole file 20 consecutive runs: 20/20 pass; test_osplat.py passes; ruff check clean; gitleaks 0 findings.

🤖 Generated with Claude Code

[nt-nerdtechnic]

Squash-merged into main as 11178c3.