Feed bitwise collectors straight into the histogram; cut reduce allocations

[MauroToscano]

Final follow-up from the #786 review, stacked on #792. Applies the PR's own "never materialize a Vec whose only consumer is the multiplicity count" principle to the remaining big sources, and removes the rayon reduce-identity allocations. Multiplicities are unchanged (same counts, same cells).

Changes

• Collectors now bump a BitwiseHistogram directly (Fn(&mut BitwiseHistogram)) instead of returning Vec<BitwiseOperation>:
  • MEMW_R bumps one IS_HALFWORD per register row — previously an ~4 B/op vector over tens of millions of rows existed just to be counted.
  • PAGE bumps one ARE_BYTES per byte of every touched page — previously a full op vector per page sweep.
  • CPU padding collapses to two bump_n calls (3n ARE_BYTES[0,0] + 4n IS_HALF[0,0]) instead of an O(padding_rows) vector of identical zero ops (which also over-reserved 2×).
  • Small sources still use their pure collect_* functions; the closure folds the transient Vec in and drops it.
• reduce_with replaces reduce(identity, ..) in the rayon tree-reduce: no more zeroed 80 MiB identity histogram per reduce leaf (up to ~16 × 80 MiB of pure alloc_zeroed + merge traffic on many-core machines).
• RegRow.address shrinks to u16 (register index 0..=255): the walk's largest persisted array goes from 40 to 32 B/row.

Measurements

Local tests

cargo build --release --features instruments, cli prove --time, Apple Silicon, best of 2 runs, this branch vs its base (#792):

workload	p4_bitwise_collect before	after	Δ
fib_iterative_1M	112.0 ms	97.8 ms	−13%
fib_iterative_4M	131.1 ms	83.3 ms	−36%

prove_total and peak RSS unchanged within run-to-run noise on this machine (p4 is <1% of total here, and macOS RSS jitters ~1.5 GB between runs). The reduce-identity savings scale with core count, so the effect should be larger on the many-core provers #786 targets. The remaining serial base.add_ops over the in-walk lookups was left as-is: parallelizing it adds code for a phase already down to ~85 ms.

Server

Testing

• Full cargo test -p lambda-vm-prover --lib on this change: 503 passed (same 5 known missing-fixture failures, which fail at file-read).
• make lint passes; clippy --all-targets no warnings; fmt clean.

[MauroToscano]

Rebased and composed with 3ba4aac: kept its par_chunks cap (peak memory no longer scales with core count), and layered this PR's two changes on top — collectors bump the chunk histogram directly (no per-source op vectors for MEMW_R / PAGE / padding), and reduce_with removes the remaining per-leaf 80 MiB identity allocations that reduce(identity, ..) still paid. Full suite re-run on the rebased stack: 502 passed (one test fewer than before — the flaky privacy byte-scan test was removed upstream in #789), same 5 known missing-fixture failures.

[MauroToscano]

/bench

[github-actions]

Benchmark — ethrex 20 transfers (median of 3)

_{Table parallelism: auto (cores / 3)}

Metric	main	PR	Δ
Peak heap	72964 MB	72876 MB	-88 MB (-0.1%) ⚪
Prove time	57.860s	38.595s	-19.265s (-33.3%) 🟢

🎉 Improvement detected — heap or time decreased by more than 5%.

✅ Low variance (time: 1.6%, heap: 1.4%)

_{Commit: 47948a2 · Baseline: cached · Runner: self-hosted bench}