parity runner: node-oracle helpers leak under fork pressure until the process table saturates (unkillable ?E state, reboot required)

[proggeramlug]

During a long session on 2026-07-25, repeated ./scripts/run_gap_tests.sh runs accumulated hundreds of orphaned node-oracle helper processes (perry_parity_nod*, comm truncated), reparented to PID 1 in state ?/?E — unkillable (kill -9 dispatched per-PID, no reap) — until the per-user process table saturated: fork: Resource temporarily unavailable from the gap runner itself, then from every shell (ps hanging, git push failing). Only a reboot cleared them.

A single post-reboot gap run leaked nothing (3 helpers before and after), so the leak likely needs load/fork-pressure conditions — e.g. the runner failing to reap oracle processes when its own forks start failing, which then compounds.

Suggested hardening: the runner should (a) waitpid/kill its oracle helpers in an EXIT trap per test, (b) run helpers in a process group and signal the group, (c) bail out early when fork starts failing instead of spawning further.

Evidence: process-table snapshot showed dozens of (perry_parity_nod) entries with PPID 1; count check between runs: ps -u $(id -u) | grep -c perry_parity_nod.

[proggeramlug]

Verified fixed on current main by the containment merged in #6762, with the later EXIT-path sweep from #7436 covering interrupted/resumed runs as well.

The process-name evidence in this issue points at compiled Perry test binaries, not Node itself: perry_parity_nod… is the run-scoped executable name perry_parity_<test-id> for tests whose IDs begin with node. That matters because #6762's cleanup is deliberately scoped to those executables under the parity run's PARITY_TMP.

Bounded reproduction on current main:

• Control: a fake compiled Perry test spawned one child and exited. The child remained alive with PPID 1, reproducing the underlying orphan lifecycle without creating process pressure.
• Canonical runner: the same fake compiler/test was run through run_parity_tests.sh --filter 1831_yield_star_array, which is the runner scripts/run_gap_tests.sh delegates to. The harness exited 1 only because the fake binary intentionally produced an output mismatch; its orphan child was reaped before the harness returned.
• No fixture process remained afterward.

The post-issue fixes now provide both layers this ticket requested: a per-test process cap plus run-scoped orphan sweep after every Perry invocation (#6762), and the same run-scoped sweep in the EXIT cleanup path (#7436). Closing as completed rather than duplicating those changes.