bench: use BatchSize::PerIteration for cold-cache variants

[max-sixty]

Summary

Cold benchmark variants in list.rs, remove.rs, and time_to_first_output.rs were passing BatchSize::SmallInput to iter_batched. Under SmallInput, criterion calls the setup closure once per batch up front and then runs the routines back-to-back inside a single timing window — so when setup is invalidate_caches_auto, only iter 1 per batch is actually cold and iters 2-N read from the cache the previous routine just populated. The reported "cold" medians were warm-biased averages.

BatchSize::PerIteration switches to setup → time(routine) per iter, so every measured iter is genuinely cold. The setup is far cheaper than a wt subprocess, so per-iter Instant::now overhead doesn't dominate. Codex flagged this on the sibling picker_preview bench (#2721); that PR landed PerIteration for its cold path, and this aligns the rest of the bench suite.

Measured corrections

Variant	Before median	After median	Spread (before → after)	Median Δ
list full/cold/8	113.5 ms	107.1 ms	16.0 → 3.9 ms	−5.7% (n.s., p=0.14)
remove_e2e/first_output	48.2 ms	86.4 ms	2.6 → 17.9 ms	+44.4% (p<0.001)
first_output/remove	50.5 ms	69.6 ms	2.4 → 0.65 ms	+23.0% (p<0.001)

remove_e2e/first_output is the starkest correction — compute_integration_lazy writes is-ancestor / has-added-changes / merge-add-probe entries, so warm-bias was substantial. first_output/remove is the cleanest demonstration of variance tightening (3.7× tighter spread). list full/cold/8 doesn't move the median significantly but the upper-tail outliers that warm-bias was inflating disappear.

Scope

Only the four iter_batched(… invalidate_caches_auto, …, BatchSize::SmallInput) call sites are switched:

• benches/list.rs run_benchmark (covers skeleton, full, worktree_scaling, many_branches, divergent_branches cold variants)
• benches/list.rs bench_real_repo cold path
• benches/remove.rs first_output
• benches/time_to_first_output.rs remove

Other BatchSize::SmallInput sites in remove.rs (no_hooks, with_hooks) use recreate_worktree as setup — that doesn't invalidate a cache the routine repopulates, so they stay as-is.

benches/CLAUDE.md "Cache Handling" gets a paragraph so future bench authors reach for PerIteration when the setup invalidates a cache the routine repopulates.

Test plan

• cargo run -- hook pre-merge --yes — 3667 tests pass, 0 failed
• Each cold variant run before+after on this branch with --save-baseline / --baseline