[plotarmordev]

Summary

Two more tests that failed only under heavy machine load. Both were test problems, not product bugs. Test-only change, apart from making one function visible to its test module.

• process::tests::cancellation_interrupts_normal_exit_pipe_drain
  • Cause: it raised a real process-wide SIGINT. A concurrently running test could install its own interrupt handler and receive that signal instead. Reproduced 5/40 with both tests running together.
  • Fix: the test now calls the capture guard's handler directly; cancel is pub(super). Real signal delivery stays covered end to end by external_program_cancellation_cleans_only_owned_group, which signals a separate process.
• systemone_tests::compare_reports_a_declared_change_under_the_observed_envelope
  • Cause: it assumed a fixed 40 ms delay always outlasts the fast server's bursts. Under load, bursts reached 1.47 s. Reproduced 18/96.
  • Fix: the fast capture runs first, and the slow server's delays follow the observed makespans.
  • New assertions: they compare the report's reference and candidate envelopes against exact extrema computed independently from the retained A+A2 and B waves. That closes a coverage gap where omitting A2 from pooling went unnoticed.
• Not reproduced: sparkdash_portable_copies… stayed green over 540 loaded runs plus injected delays, so it's left unchanged.

Testing

• cargo fmt --check and cargo clippy --all-targets -- -D warnings.
• Loaded loops: 240/240 and 150/150 runs passed per test.
• Mutants: drain cancellation, the cancel flag, reversed gate direction, swapped pooling, omitting A, and omitting A2 are all caught.
• Review: built by WRITER, then SIX for round 2. A blind SIX review approved after one round.

Performance claim

None.