Engine Fixes - 2026-11-01

Context

The Engine and CLI Bug Hunt (2026-10-31) fixed the small bugs it found and left 15 proposals (P1 to P16, no P15) that change behaviour or are bigger than a one-line fix. Taltal approved all of them on 2026-11-01. They ship as several small pull requests, one area each, so each can be reviewed and reverted alone.

Decision

#	Change	Where
P1	One run of a cycle at a time per artifacts folder. CycleRunService.run holds an OS file lock on <artifacts>/<cycle>/.lock (fcntl.flock, msvcrt.locking on Windows) from the first stage until the reporters, native report snapshot and report archive have read the folder. A second run of the same cycle waits and prints another run of cycle 'x' is using ... (pid N); waiting for it to finish. Other cycles, and the same cycle with another artifacts_root, never wait. The OS drops the lock when the process dies, so a killed run never leaves it held.	qasight/engine/cycle_lock.py, qasight/services/cycle_run.py
P2	Stage folders a run didn't write are removed. A stage renamed or deleted from the config left its folder (with allure-results/) behind, and the next run counted its old tests. At the end of a run the orchestrator deletes child folders of <artifacts>/<cycle>/ that this run didn't write and that hold run.log or allure-results/ (names compared case-insensitively, symlinks skipped).	qasight/engine/orchestrator.py
P3, P4	A timeout, Ctrl-C or SIGTERM stops the whole stage and records the run. Stages start in their own process group (start_new_session, CREATE_NEW_PROCESS_GROUP on Windows) and are signalled as a group, so grandchildren (npm → node, xdist workers) don't outlive a timeout. The first SIGINT/SIGTERM terminates the running stages, the second kills them; no new stage starts, the run is recorded FAILED with interrupted_by, plan_aborted has reason signal, and qasight run exits 130 (SIGINT) or 143 (SIGTERM).	qasight/engine/interrupt.py, qasight/engine/backends/__init__.py, qasight/engine/executor.py, qasight/engine/exit_codes.py
P5	Keys nothing reads are warned about. The raw config is walked against the JSON Schema; an unknown key prints a warning with a did-you-mean (env → extra_env, timeout → timeout_s), and qasight config validate --strict exits 2 on one.	qasight/config/unknown_keys.py
P6 to P8	Safe on_release, shell-style extra_env, strict if:. A pool's on_release hook runs only after a heartbeat confirms the lease is still held; extra_env turns true/false/null into true/false/empty instead of Python's True/None and refuses lists and maps; an if: outside the supported grammar is a config error at load time instead of silently true.	qasight/resources/client.py, qasight/config/if_expr.py, qasight/config/sections.py
P9, P10, P16	Clear CLI outcomes. qasight split --index exits 1 for an empty shard (--allow-empty exits 0), since an empty pytest $(qasight split ...) runs every test; qasight badge and qasight flaky name the closest recorded cycle; a --server that isn't an http(s):// URL exits 2 in split, ingest and lease.	qasight/cli/commands/split.py, qasight/cli/cycle_hint.py, qasight/cli/server_url.py
P11 to P14	Real run times, no escaping attachments, Windows Jest paths, XML whitespace. Engine runs store their real start_time/end_time; an Allure attachment symlinked outside its folder is skipped; Windows paths in Jest reports are made relative like POSIX ones; whitespace before <?xml is stripped.	qasight/persistence/db_backend.py, qasight/reporting/allure_passthrough.py, qasight/reporting/jest_import.py, qasight/reporting/junit_import.py

Each row is one pull request.

Consequences

• P1: two CI jobs of one cycle on one runner now run one after the other instead of overwriting each other's results. The API already refused a second run of a busy cycle in its own process; the lock also covers the CLI next to the API. With --async-report-db the background archive reads the folder after the lock is released, so a run that starts right then can still race it; per-run folders (artifacts/<cycle>/<run-id>/) would remove that and are left for later.
• P3/P4: 130 and 143 are new exit codes (an additive change to the frozen exit-code contract); scripts that treat any non-zero as failure are unaffected.
• P5 and P8: a config that used to load silently may now warn (P5) or fail to load (P8, a bad if:), which is the point: both hid mistakes.
• P2: a folder under artifacts/<cycle>/ that holds run.log or allure-results/ and isn't a stage of the current run is deleted when the run ends. Earlier runs keep their copy in static/history/<run_id>/; only the shared working folder is cleaned.