A signal that has been red for eleven minutes does not say whether the disk
stalled once for eleven minutes or four times, with quiet in between. Both
histories produce the same verdict if the verdict only reports "RED". We found
that the distinction was sitting in the probe records all along, and that two
small changes to the cause line made it visible without changing the alert.

A state is not a history

An alert value describes the present: is the condition true right now? A
responder needs something else — did this just start, or did it come back after
stopping? Those are different situations. A first stall is a new problem; a
fourth stall after two minutes of quiet is a problem that was never solved and
decided to recur.

Sustained alerts hide that difference structurally. Once the value is red, a
second event inside the window produces no new edge. Nothing changes state, so
nothing wakes a resolver. If the only description of the condition is "red for
eleven minutes", the recurrence is real but invisible.

What the records showed

The check keeps individual probe records, so the window can be replayed. In one
measured window spanning 08:00 to 08:11Z there were four separate bursts of
stalls, the newest ending at 08:11:51Z. They were not four readings of the same
slow operation — they were separated by genuine quiet.

We chose a simple grouping rule: a quiet gap of at least two minutes starts a
new burst. A multi-probe cliff is still one episode, because repeated readings
of the same slow operation are collapsed before the grouping runs. That detail
matters: without the collapse, a single cliff sampled three times would inflate
the burst count and turn one event into three.

The worst stall is not the peak stall

The second finding came from the same records. The verdict reported pressure
only for the stall that lasted longest. In this burst that was a 23.6-second
stall carrying an I/O pressure reading of 57. But the peak I/O pressure in the
window was 93.23, on a shorter 14.9-second stall — and the only non-zero memory
pressure, 0.22, was on that shorter stall too.

Reporting only the longest stall would have told the wrong story. The two
readings point at different causes: sustained device latency versus a reclaim
spike. A responder acting on the first would look in a different place from one
acting on the second. The data to distinguish them was already collected; it
just was not in the signal.

What changed

When a window contains two or more bursts, the cause now names how many there
were and when the newest one occurred. When a window contains two or more
stalls, the cause reports peak I/O and peak memory pressure across all of them,
not just the reading from the longest one.

Thresholds, the window length and the signal value are unchanged. A single
stall does not get a repeated peak reading, because there is nothing to compare.
The change is to the explanation attached to the verdict, not to when the
verdict fires.

Boundaries and checks

The grouping threshold is a definition, not a discovery: two minutes is a
choice about what counts as separate, and a different value would group the
same records differently. The number is stated in the cause so a reader knows
which definition they are looking at.

Focus tests cover the boundary cases — nearby episodes staying one burst, a gap
just over the threshold splitting into two, a multi-probe cliff not inflating
the count, and a single stall not repeating its pressure reading. When probe
records cannot be read, the verdict stays UNKNOWN rather than quietly reporting
a clean window.

The habit

When you read a sustained alert, ask two questions before acting. How many
separate events does it contain? And which event carried the peak, as opposed
to the one that lasted longest? The answer changes what you go and look at.
A red light is a useful summary, but a summary is not the whole record.