bench refcount growth — phase 04: fuzzer retirement + docs¶

Parent: PLAN-bench-refcount-growth.md. Planned at medium effort.

Scope¶

scripts/differential-fuzz.py: remove the qcow2-write-refblock-coverage steer-around — the _BENCH_WRITE_SAFE_CLUSTERS constant, the post-invocation cluster-size override in _bench_option_picker, and the docstring bullet — restoring the full QCOW2_CLUSTER_SIZES matrix for -w recipes. (Removing the extra rng.choice draw restores the pre-steer RNG stream, so the original issue iterations replay exactly.)
docs/bench.md: drop the qcow2-write-refblock-coverage known-divergence bullet; the "Write tests" growth text landed in phase 2.
CHANGELOG.md: Unreleased → Added entry describing the growth feature and the retired limitation.
docs/plans/index.md + order.yml: master-plan row and nav entry (if not already added when the plan landed); set the status to Complete at the end of this phase.
Validation (management session, not sub-agent):
replay iter_seeds 978924777, 2808571592, 2843295063, 933136565, 3155371439 against the rebuilt binary + updated fuzzer — all must report divergence: null;
a ≥ 250-iteration soak (fresh seed) — 0 divergences;
full test suites + pre-commit.

Step	Effort	Model	Isolation	Brief for sub-agent
4a	medium	default (Fable)	none	Apply Scope items 1-4. The steer-around was added in commit ee2cc5d — use it as the reverse map, but keep the KNOWN_BENCH_DIVERGENCES history accurate: the docs/bench.md "Known divergences" list simply loses the bullet (registry entry removal is phase 3's). Keep the CHANGELOG entry factual: what grew, the new refusal envelope (needed slots > 2048 / staged bytes > 2 MiB / RT > 64 KiB), the crash-window note, and the fuzzer-matrix restoration.

Standard checklist, plus the parent plan's success criteria (issue replays + soak) executed and recorded before marking the master plan Complete.