F1: Drop BP's rank_l2 (largest remaining select/rank memory win)

[newhoggy]

Status (rescoped 2026-08-08, updated 2026-08-08)

Unblocked — #665's rank1 micro-benchmark merged. benches/bp_rank_micro.rs now covers BalancedParens::rank1 across flat/deep/mixed tree shapes at 1K-1M opens, benchmark-only (no src/ changes). Pre-#596 baseline (10,000 rank1() queries/batch, µs):

Platform	1K	10K	100K	1M
ARM (M4 Pro)	~11.5	~11.6	~11.9	~12.8
x86_64 (Zen 4 7950X)	~26.7	~24.4	~24.7	~27.9

Shape (flat/deep/mixed) is neutral on both platforms — expected, since the current rank_l2 decode is a pure shift-and-mask independent of bit content; size drives the difference, with 1M falling out of cache on both architectures. This is the number any #596 implementation attempt should diff against on both rank1 (via this bench) and select1 (via the existing benches/bp_select_micro.rs, from #64) — both consumers still apply, this issue is otherwise unchanged from the assessment below.

---

Summary

rank_l2: Vec<u64> in BalancedParens costs 64 bits per 512-bit block =
12.5% of the BP bitmap — twice what #64's Step B saved. Poppy's actual
trade is to omit the per-word level and popcount up to 7 words inside a basic
block, which is exactly one 64-byte cache line and one the query must load
anyway for the partial word.

Why now

Interacts with #64's Step B, but not the way this issue originally stated.

The claim was that once select drove off rank_l1, rank_l2 would be left
with rank1 as its only consumer. That is wrong as #64 was implemented:
WithCsPoppy::select1 reads rank_l2 too, unpacking its 9-bit per-word
offsets (step 3 of its query) to land on the exact word without popcounting.
rank_l2 now has two consumers, not one.

That makes this issue more expensive than "delete an array", not less.
Dropping rank_l2 forces select1 back to popcounting up to 7 words inside
the block — reintroducing exactly the word scanning Step B removed, and most
likely widening the ARM select1 regression #64 measured (+6-14% micro, +7.4%
end-to-end on M4 Pro). Budget for changing both query paths, and re-run
bp_select_micro alongside the rank-side measurement rather than treating this
as rank-only work.

BP rank1 is very hot (cursor navigation), so the rank side remains a genuine
space/time trade needing end-to-end measurement, not a free reduction. Should
follow #64 so the select path isn't confounded with this change.

Deferred from

#64 (docs/plan/cspoppy.md §6, F1) — not scheduled, recorded so the reasoning
isn't lost.