[bench][retention] Exclude a cell the agent never finished

[bherrera]

Summary

A cell whose agent exited non-zero was scored as a task failure. The code
already said it should not be —

# A cell that failed for a reason outside the thing being measured has
# to say so. Silently scoring it as a task failure would blame the
# filter for the harness.

— and then set a subtype that nothing read. attempted() asked only whether
tokens had been spent, and a run cut off partway has spent some.

What it cost

A Sonnet sweep was blocked partway. It recorded this:

cargo-cascade level1 #3   13,493 tokens   2 tools   pass=False
   note: agent exit 1:

and reported knee: success falls below 100% at level1. There was no knee.
The agent was cut off by the same block that left the other 51 cells
unattempted; it had simply spent tokens first.

Rescoring the same recorded cells with the guard wired:

	knee
cut-off cell counted	level1
cut-off cell excluded	None

Test plan

ruff check bench/runner.py

Verified against the recorded run above, plus the three cases directly: a
finished cell still counts, one that never started still does not, and one cut
off mid-run is now excluded and named in the report.

Note for reviewers

This is the second guard in this file written and left unwired — cell_note()
was the first. The pattern is worth watching for: the comment states the
intention, the code sets a field, and nothing consumes it. A benchmark's guards
are exactly the code that never runs on a good day.

The consequence was not cosmetic. --save-baseline refuses a curve with
unattempted cells, but this cell counted as attempted, so a blocked sweep could
have been saved as a baseline asserting a knee that does not exist.

No curve is committed here. The sweep that exposed this got 9 of 60 cells, which
is not a curve, and the Sonnet baseline still predates the level 0 change.