[thebhef]

Fault-recovery bench: six PEAK runs of the shipped recovery recorded

Trials 7–12 of 2026-10-06, close-then-open shipped as the sidecar's own
bus-off recovery (strategy sidecar): bus-off recovered 3/3 (trials 8,
10, 12 — twice at the plug, once at the pull), error-passive recovered by
itself 3/3 (7, 9, 11), the stuck transmit queue flushed and resumed (7, 8).
Run 7's 1.8 idle hours before the pull are trimmed from its events file.

Co-Authored-By: Claude Fable 5.1 noreply@anthropic.com

Bench: adapter-clock timeline and a bus_back event; ADR 0039 close-before-open evidence

Phase 9d. The bench now tracks when each channel's own bus-error
episode and the pull/bus-off/reopen/recovered stages happened on the
adapter's own clock, not just the bench's software detection: a
bus_back event fires once the channel under test reads active with
its episode closed and hears the partner again -- a reopened channel
has no backlog, so this is close to live, unlike recovered, which
waits out whatever backlog the partner is still draining. The verdict
now prints and records this timeline (table.md and the verdict event),
plus how far behind wall clock the slower channel's last error report
arrived at bus back.

Fixed along the way: the timeline's own "recovered" entry disagreed
with the verdict's by about a second, because recovered is confirmed
only after its clean run has held for recovery_window_s -- back-dating
the wall-clock reading to when the run actually started (what
run_start_t reports) rather than to the confirming frame's arrival
fixes it.

ADR 0039's 2026-10-06 amendment gains the two findings the live PEAK
runs established but the amendment did not yet say: a pull alone does
not take a controller bus-off (an error-passive transmitter's
acknowledge errors do not raise its TEC; bus-off came only from
bit/form/stuff errors at the connector), and the measured recovery
time (active within about 1.6 s of bus-off), with the strategy-ladder
verdict table as evidence. ADR 0060's pointer to ADR 0039 already
covered close-before-open and the bus-off-refusal re-arm, so it is
unchanged.

Comment fix: driver_python_can.py's _BUS_OFF_TEXTS block and
bench/fake.py's open-failure comment both named "the owner's bench" as
the source of a claim; both now state the PCAN-Basic rule on its own
text, citing ADR 0039 where fake.py needs a reason.

Sidecar README's Fault recovery bench section documents the bus_back
event, the timeline, that wait's own --timeout is separate from run's,
and a short paragraph on what the strategy ladder showed on PEAK.

Checks (sidecar-scoped; see report): ruff check/format, mypy, full
pytest (372 passed) all clean; comment-references grep clean.

[thebhef]

Merge activity

• Oct 22, 5:42 PM UTC: A user started a stack merge that includes this pull request via Graphite.
• Oct 22, 7:02 PM UTC: Graphite couldn't merge this pull request because a downstack PR Fix the host README's stale stuck-queue paragraph for ADR 0060 #550 failed to merge.