Add readiness check to tests to avoid race condition flake

[asmacdo]

Three tests assumed duct takes at least one sample of a short command (sleep 0.1 / sleep 0.3). On a slow runner (one ps sample takes up to 0.72 s on macos-15-intel) that is not guaranteed, so they failed with zero samples. They now keep the command alive until its pid appears in duct's usage file, so no test depends on timing.

Together with #456 this fixes the nightly Intel Mac failures; any nightly failure after both merge gets a new issue.

Closes #451
Closes #439

Stacked on #456: this branch is based on it, so the diff shows #456's commits too until #456 merges. Only the last 5 commits are this PR.

Changes

• New test/data/until_sampled.sh USAGE_FILE: loops until its own pid shows up in the usage file (gives up with exit 1 after 60 s).
• test_session_modes: runs until_sampled.sh instead of sleep 0.3; exit code 0 also proves the command was sampled.
• test_execution_summary: same, with a 0.1 s report interval; rerun backoff dropped.
• test_spawn_children: each child records its pid in a test-provided dir and stays alive until sampled; the parent script exits only once all were sampled (replacing the hold-the-door sleep, so the expected count is N, not N+1). The test compares recorded pids, not command-line text (which ps -w truncates), and requires every child to have recorded itself.
• test_spawn_children, setsid case: the process that calls setsid now forks the recording child, so that child was never in duct's session and "its pid must not appear" is a fair check. The script waits for every child's record before counting two more reports. The case is skipped where there is no setsid command.
• The flaky rerun markers on test_spawn_children, test_signal_int, test_signal_kill and test_execution_summary are removed; no test reruns on failure any more, so pytest-rerunfailures is dropped from tox and the marker's registration from pyproject.toml.

Verification

Local, Linux, py3.13, pytest-rerunfailures disabled, whole process tree capped with systemd-run --user --scope -p CPUQuota=10% (measured before the setsid change):

• spawn_children + session_modes: 84/84 over 3 passes (before: 25 of 28 failed).
• test_execution_summary: old 8/8 failed, new 5/5 passed.

The setsid change: all 24 spawn_children cases pass over 5 passes, and also with setsid delayed by 30, 50 and 200 ms through a PATH shim. Those delays made the previous version fail 7 of 20 and 20 of 20 times, or pass with no children recorded.

CI is the real check, especially the Intel 3.10 job, which reran these tests 7 times on #456. On this branch every job passes with zero reruns. The macOS runners have no setsid command, so the setsid cases are skipped there (they used to pass without checking anything) and run on Linux.

Why the setsid case needed fixing

An earlier push of this branch failed test_spawn_children[duct-1-setsid] on 2 of 36 Linux cases: a recorded child pid showed up in duct's samples. setsid cmd & forks inside duct's session, and that same pid only then execs setsid, calls setsid(2) and becomes the child. For those few milliseconds the pid really is in the session duct watches, so a sample landing there recorded it, and the test compares pids over the whole run. The old test did not notice because it matched on sleep in the command line.

Why one sample is slow on Intel Mac

One duct sample is ps -ax plus a session lookup per process (~500 processes on runners). Measured on CI with a diagnostics branch (single runs, no spread):

runner	one sample (median / max)	usage lines for sleep 0.3
Linux	0.004-0.009 s / ≤0.08 s	7
ARM Mac	0.03-0.06 s / ≤0.11 s	2-5
Intel Mac	0.05-0.24 s / up to 0.72 s	1-4 (3.10: 1,1,2)

test_session_modes[current-session] never failed: that session includes pytest and the shell, so a sample is never empty.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.27586% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 94.38%. Comparing base (7a25f42) to head (86612a2).
⚠️ Report is 7 commits behind head on main.

Files with missing lines	Patch %	Lines
src/con_duct/_signals.py	97.72%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #457      +/-   ##
==========================================
+ Coverage   92.03%   94.38%   +2.34%     
==========================================
  Files          15       15              
  Lines        1344     1389      +45     
  Branches      182      186       +4     
==========================================
+ Hits         1237     1311      +74     
+ Misses         71       48      -23     
+ Partials       36       30       -6     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

🔵 Needs a closer look

It modifies signal-handling and concurrency-sensitive production code and rewrites process/session-sensitive test scripts whose correctness depends on subtle runtime behavior best confirmed by human review and CI.

0 open findings

What changed in this PR

This PR makes three e2e/signal tests deterministic to eliminate the nightly Intel-Mac CI flakes that occurred when a short command (sleep 0.1/0.3) finished before duct took a single resource sample. Instead of relying on timing, commands now stay alive until their pid appears in duct's usage file, and rerun-on-failure machinery is removed. It is stacked on #456, so the diff also includes that PR's signal-handling source and test changes.

Changes:

• New test/data/until_sampled.sh and a rewritten spawn_children.sh that keep the command/children alive until duct has sampled them (each child records its pid, then execs the wait loop), with a bounded 60 s timeout.
• test_execution_summary, test_session_modes, and test_spawn_children updated to assert on recorded pids/samples rather than timing; setsid cases are skipped where no setsid command exists.
• pytest-rerunfailures and the flaky marker are dropped (tox deps + pyproject.toml), replaced by the stacked signal-handling fixes (_signals.py, _duct_main.py, _output.py) and their tests.

File	Description
test/​data/​until_sampled.sh	New script that blocks until its own pid is recorded in the usage file.
test/​data/​spawn_children.sh	Rewritten to record child pids and wait for sampling; setsid children forked after setsid.
test/​duct_main/​test_e2e.py	test_spawn_children/test_session_modes assert on recorded pids; adds subprocess-based signal tests (#456).
test/​duct_main/​test_execution.py	Uses until_sampled.sh; adds test_fail_time_threshold; removes timing-based signal tests.
test/​utils.py	Adds start_duct/rest_of_stderr readiness helpers and DUCT_READY regex (#456).
test/​data/​signal_ignorer.py	Reports progress to a file so tests wait on steps instead of sleeping (#456).
src/​con_duct/​_signals.py	Deferred-attach SIGINT handler, start_thread, signal-safe _log fallback (#456).
src/​con_duct/​_duct_main.py	Installs handler before command start; uses start_thread for monitoring threads (#456).
src/​con_duct/​_output.py	Starts tail thread via start_thread so it never receives SIGINT (#456).
tox.ini /​ pyproject.toml	Drop pytest-rerunfailures/flaky marker; add coverage>=7.10 and patch config (#456).
.gitignore	Ignore parallel-coverage .coverage.* files.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.