[Nir-Az]

Overview: On the D585S, the test wrappers switched safety mode immediately, but the FW can transiently reject a safety_mode change (RuntimeError: Failed to set the option to value N) — especially the first attempt after enumeration / disruptive ops under CI load. This flaked the d500 safety suite and tests entering service mode via test_device_wrapped (e.g. test_advanced_mode_support, test_depth_laser_on), bench-to-bench on identical FW.

Root cause: Transient FW readiness. A first set_option attempt often fails (blocks ~3s, throws); a retry succeeds. Across 330 logged transitions the success envelope is tight — 1 attempt (<2s) or 2 attempts (~3.6s); no transition ever succeeded beyond ~3.6s. get_safety_interface_config() isn't a reliable readiness signal, so retrying the actual set_option is what works.

Fix: tests_wrapper.set_safety_mode(safety_sensor, mode, timeout=8, interval=0.5) retry helper (uses rspy.stopwatch.Stopwatch, logs safety_mode set to <mode> after N attempt(s), Xs), routed through every D585S safety-mode transition:

• rspy/tests_wrapper.py → start_wrapper / stop_wrapper
• conftest.py → test_device_wrapped (enter + teardown)
• pytest-operational-mode.py, pytest-operational-mode-stress.py, pytest-interface-config-get-set.py (post-reboot)

Timeout is 8s (~2× the 3.6s worst observed success); larger values were tried but never recovered anything a small retry didn't — the rare truly-stuck transition recovers only via a full test rerun, tracked separately in RSDEV-12512 (module-scoped test_device_wrapped can't re-setup on pytest-retry → hard error). D585S-only; other devices untouched.

CI verification (vtglnx163 + D585S 333622320169):

• update expected filename on fw update test #14451 (-r d500 --repeat 3, 261 tests): 0 retries.
• migrate s3 links #14455 (full suite, 846 tests): 0 retries, 0 failures.
• Final D585S-only full nightly at 8s: in progress.

Jira: RSDEV-12509 (fix), RSDEV-12512 (fixture-retry fragility), RSDEV-12232 (FW-side, closed)

🤖 Generated by AI

[sysrsbuild-gh-agentic]

[rs-agentic bot] Automated review of PR #15329 — synchronize event. No Copilot reviewer detected, proceeding with full inline review.

[Nir-Az]

Thanks for the review (rs-agentic bot). Addressed in 1e2e5f902, with a few respectful push-backs:

Fixed (1e2e5f902):

• Trailing newline added at EOF.
• Stale last_exc: it's now reset each iteration, with a log.w(...) on a set-accepted-but-read-back-mismatch, so the error surfaced at timeout is never stale/irrelevant.
• Redundant teardown import in conftest.py removed (it's already bound from the setup block across the yield).

Not changing — with rationale:

• supports() guard before get_info(name) (start/stop_wrapper): this call predates the PR — the wrappers already used it; the PR only inserts the retry call. They run solely for live D585S device tests, where camera_info.name is always present, so it's never been an issue. Can be a separate cleanup if the team wants parity with conftest.
• PEP-8 spacing: the spaced style matches this file's existing convention — def start_wrapper( dev = None ): / def stop_wrapper( dev = None ): directly below use the same spacing, so tightening only the new function would make it inconsistent with its neighbors. The repo is mixed here (e.g. rspy/devices.py has both time.sleep( 1 ) and time.sleep(1)). Keeping local consistency; a file-wide PEP-8 pass can be separate.
• Stress-test retry masking regressions: two notes — (1) pytest-operational-mode-stress.py is currently @pytest.mark.skip ("disabled until HKR FW stable"), so it isn't running; (2) CI timing shows mid-session SERVICE transitions transiently need a 2nd attempt (~3.6s) even warm, so the suggested timeout=2–3s sits below the normal transient and would cause false failures. If we want fail-fast sensitivity there, ~5s is the right floor (above 3.6s, below the 8s default). Left on the 8s default for now since it's disabled; can set timeout=5 on the in-loop calls when it's re-enabled.

Already resolved: the 60s-timeout thread (reduced to 8s in c49adc8, which the bot approved) and the stale PR-description note (description updated to 8s + full-suite verification).

🤖 Generated by AI