[christianparpart]

Closes #1739.

Why

The dead-peers never-accepting leg holds the launcher's reported cache time to a ceiling of (2N - 1/2) x TOTAL, which is 7000 ms. On loaded Windows runners it reported 7602 ms (in #1737) and 8132 ms (on master). 8132 is past the 8000 ms that a launcher which doubled its deadline reports, so time alone cannot tell a stalled host from a real regression. Raising the ceiling would stop it catching a doubled deadline.

What

• A starvation witness runs beside the leg in both fixtures that hold this ceiling:
  • compile-cache-e2e.sh (POSIX) uses a perl ticker;
  • run-launcher-e2e.ps1 (which is what ctest runs on Windows) uses a Stopwatch ticker on its own runspace.
• The witness counts only stalls. It requests 20 ms sleeps and sums only the part of each wake-up that was more than 50 ms late. Raw lateness measures the timer instead: on an idle host, Git Bash's perl is late by 2.5 s over 8 s and pwsh by about 3 s, either of which would excuse a doubled deadline. The stalled figure measured 0 under perl, WSL and pwsh.
• It watches only the cache phase. The witness stops at the launcher's fall-back line, so a stall during the compile that follows is not counted.
• The verdict is a pure function with three outcomes, e2e_dead_ceiling_verdict in bash and Get-E2EDeadCeilingVerdict in pwsh:
  • pass: the reported time is within the ceiling;
  • inconclusive: the host stalled at least as long as the overrun. This is printed, raised as a ::warning:: annotation, and repeated at the end of the run. The run does not fail;
  • fail: the host did not stall that long, or the witness left no reading. A missing reading fails closed.
• Self-tests: the bash verdict has cases in e2e-helpers-selftest, and the pwsh verdict and witness have the new launcher-e2e-starvation-witness-selftest. A pwsh case asserts that the 20 ms and 50 ms constants match across the two languages.

Blind spots (stated in both files)

• A host that starves the launcher but not the witness fails OPEN.
• Starvation in slices shorter than 50 ms counts nothing, so it fails CLOSED. The noise is reduced, not eliminated.
• ctest shows output only on failure, so an inconclusive warning on a passing run reaches GitHub only from the jobs that run the fixtures directly.

Tested

• Self-tests:
  • bash: check-e2e-helpers.sh ran 1147 checks, 0 failed (WSL);
  • pwsh: the witness self-test ran 32 cases, 0 failed, and -SelfTestPorts ran 47.
• Each guard bites: each was disabled in turn, and only its own case went red.
• End-to-end: run-launcher-e2e.ps1 ran on Windows (cl and clang-cl). The never-accepting leg passed with 0 ms stalled, and the witness closed at the end of the cache phase.