perf(prover): pull-cursor CPU scheduler instead of fixed chunks, plus…

[diegokingston]

… tests

Two follow-ups to the CPU-aware scheduler.

Keep the pull cursor. The CPU run_admitted batched order into chunks(workers) and ran each chunk to completion, so every table waited for the slowest member of its chunk — the behaviour the doc comment directly above run_admitted says heaviest_first plus the cursor were introduced to remove.

Sorting heaviest-first hides this while the weight estimate is accurate: chunks then hold similar-weight work. It bites when the estimate mispredicts, which is plausible here since the weight is estimate_table_vram_bytes — a device footprint used as a proxy for host time.

This keeps the original CUDA scheduler's shape (atomic cursor, heaviest-first, concurrency bounded by workers) but spawns the workers as Rayon tasks rather than OS threads. That preserves the fix this branch is for — nested per-table Rayon work stays inside the pool and work-steals — while a worker that finishes a small table pulls the next one immediately. TABLE_PARALLELISM keeps its meaning and the number of simultaneously live per-table working sets is unchanged. Blocking a pool worker is only acceptable because, unlike the CUDA path, there is no admission gate to wait on.

Add coverage. run_admitted is now three cfg-gated implementations with no tests. These are cfg-agnostic, so they exercise whichever one the build selects:

• results land in the slot named by their own index, not by position in order (those differ once the order is heaviest-first);
• a batch does not serialize when the estimate mispredicts — verified to FAIL at 565ms against the chunked implementation and pass at ~130ms here;
• heaviest_first is a descending permutation.

Not addressed here, to keep the diff reviewable: the CPU staging block is a near-duplicate of the debug-checks one, and VramGate is still constructed and threaded as _gate on CPU rather than not built.

[jotabulacios]

/bench

[github-actions]

Benchmark — real block (ethrex_mainnet_25368371.bin) (median of 3)

_{continuations · epoch 2^22 · 12 epochs}

Metric	main	PR	Δ
Peak heap	47024 MB	50014 MB	+2990 MB (+6.4%) 🔴
Prove time	157.521s	142.752s	-14.769s (-9.4%) 🟢

❓ -9.4% — beyond what 3 runs resolve. Use /bench-abba for a paired test of the same block (default 12 pairs, ~72 min, resolves ~1%).

_{Prove-time spread 0.5% (142.929s / 142.752s / 142.221s)}

_{Commit: 0f2f293 · Baseline: cached · Runner: self-hosted bench}