test: harden stress_no_fd_leak against shared-runner fd churn

[enricopiovesan]

Summary

Harden the flaky stress_no_fd_leak stress test so it no longer fails on unrelated descriptor churn on shared CI runners (it failed on docs-only PR #572 on x86_64-unknown-linux-gnu, then passed on re-run), while keeping it a meaningful leak detector.

Governing Spec

• 047-thread-pool-executor

Acceptance scenario 5.2 of spec 047 ("Given repeated executor create-use-drop cycles, When resource checks run, Then no file descriptor or handle leak is detected") still holds: the hardened detector fires on any genuine per-cycle leak (verified by injecting one leaked descriptor per cycle, which fails the test).

Project Item

• Closes Flaky stress_no_fd_leak on shared CI runners: absolute fd-count assertion is noise-sensitive #574 (on https://github.com/orgs/traverse-framework/projects/1/)

What Changed

• Contracts changed: none.
• Runtime behavior changed: none — test-only change in crates/traverse-runtime/tests/thread_pool_stress.rs.
• Compatibility impact: none.
• ADR needed or linked: no.

Test hardening details:

• 50 warm-up create-use-drop cycles run before the baseline is captured, so lazy process-wide initialization (stdio, threading, platform probes) opens its descriptors outside the measured window.
• Each measurement is the minimum of 5 samples 100ms apart, filtering descriptors that unrelated runner activity opens transiently.
• The assertion is growth-based with a fixed tolerance of 8 descriptors over 1,000 measured cycles, replacing the noise-sensitive absolute after <= before check. A real per-cycle leak produces growth of ~1,000 descriptors, so the tolerance cannot mask one.
• Up to 3 re-measurements (500ms apart) run before the assertion fails, letting transient churn settle.

Definition of Done

• stress_no_fd_leak uses warm-up baseline + growth-with-tolerance assertion + retried measurement
• Any real per-cycle fd leak still fails the test (validated by temporarily injecting std::mem::forget(File::open(...)) per cycle — test failed as expected)
• Full stress test suite passes locally

Validation

• Spec alignment checked (BASE_SHA=origin/main bash scripts/ci/spec_alignment_check.sh on this body)
• Contract alignment checked (no contract surface touched)
• Tests updated and passing (cargo test -p traverse-runtime --test thread_pool_stress -- --ignored --nocapture: 9 passed, 0 failed)
• Core coverage preserved (test-only change; no runtime code touched)
• Required validation gates passing
• cargo fmt --check and cargo clippy -p traverse-runtime --tests clean
• Leak-detector sensitivity re-verified: injected 1 fd leak per cycle → test fails; removed → test passes

Notes

Fixes the flake reported in #574. The tolerance (8 fds) is deliberately small relative to the 1,000-cycle window so the test cannot be silenced by threshold inflation — any per-cycle leak exceeds it by two orders of magnitude.