docs: correct a false claim, perf does sample on the bench (#452)

[jdatcmd]

design/ISSUE_452_ROUTE_B_VS_PHASE_1.md on main states that perf cannot sample
on the bench box, and uses that to justify replacing the symbol-bucket
attribution with a differential design. The claim is false.

The probe behind it was:

perf record -e cpu-clock -- sleep 1   ->  file written, "has no samples!"

sleep consumes no CPU, so a cpu-clock profile of it has no samples whether or
not the profiler works. The probe could not produce the outcome it was testing
for, so it could not distinguish the two cases.

What is true is narrower: this host has no hardware PMU, so cycles is genuinely
unsupported, which is what #505 records. Software events work. Extending the
first fact to "cannot sample at all" was the error.

Re-checked before opening this, on jd-laptop rather than the bench, because the
point at issue is whether the probe can distinguish the two outcomes:

probe	event	samples
sleep 1, the original probe	cpu-clock	2
CPU-burning loop, same box, same event	cpu-clock	4K, 0 lost

The near-empty first result reports an idle workload, not a broken profiler. The
commit records the equivalent re-run on the bench itself: 1K samples, zero lost.

The differential method is not withdrawn. It measures the milliseconds each route
would remove rather than inferring a saving from symbol shares, which is better
evidence for the decision either way. What changes is the record of why it was
chosen, and the note that the attribution it replaced was available.

Documentation only. No code, no test, no build surface.

The commit is bdaacda from feat/452-late-materialization, an unmerged local
branch found during a branch sweep. It is cherry-picked onto current main here
and applies cleanly. Every other branch in that sweep was verified already
contained in main and deleted.

🤖 Generated with Claude Code

https://claude.ai/code/session_01CmUVgsuNiCxuf3FjQKFoGx