feat(prover): disk-spill mmap backend for large program proving

[ColoCarletti]

Summary

Adds mmap-based disk spilling for trace tables, LDE columns, and Merkle tree nodes, enabling proving of large programs
(64M+ instructions) that would otherwise OOM.

Changes

Crypto layer — mmap backing for core data structures:

• Table<F>: spill_to_disk() moves trace data from heap to mmap'd tempfile, get() reads transparently from either
  source
• LDETraceTable: spill_main_from_pool() and add_aux_from_pool() persist LDE columns to disk before pool buffers are
  reused
• MerkleTree: spill_nodes_to_disk() moves tree nodes to mmap, accessed via node_get()
• FieldElement<F>: added #[repr(transparent)] for safe mmap pointer cast
• Fingerprint optimization: avoid redundant ×1/×(-1) multiplications in LogUp

Prover pipeline — chunked processing with disk spill:

• Phase A (main commit): spill main trace → extract to pool → LDE → spill LDE + Merkle tree → reuse pool for next chunk
• Phase C (aux commit): build + commit K aux traces at a time, spill each, free heap
• Phase D (deep composition): read LDE directly from spilled mmap — no recomputation

Trace builder — memory-bounded bitwise processing:

• chunk_generate_and_spill: generates each trace chunk and immediately spills to disk
• Incremental bitwise: flush_bitwise! macro updates multiplicities in 1M-row batches instead of accumulating all
  bitwise ops (saves ~14 GB at 64M instructions)
• Early drops: memw_ops, lt_ops, cpu_ops freed as soon as their traces are generated

CLI:

• prove and verify commands with --max-rows flag for per-table chunk size control
• Disk-spill status indicator

All changes are behind the disk-spill feature flag (default for CLI). Without the flag, behavior is identical to main.

Test Results

Program	Instructions	Prove + Verify
sub	6	✅
fib_iterative_2M	2M	✅
fib_iterative_32M	32M	✅ (server)
fib_iterative_64M	64M	✅ (server, 128GB RAM)
fib_iterative_128M	128M	OOM (raw ops exceed 128GB before disk-spill kicks in)

[ColoCarletti]

/bench

[github-actions]

Benchmark — fib_iterative_8M (median of 3)

_{Table parallelism: 32 (auto = cores / 3)}

Metric	main	PR	Δ
Peak heap	256381 MB	230543 MB	-25838 MB (-10.1%) 🟢
Prove time	99.470s	293.881s	+194.411s (+195.4%) 🔴

⚠️ Regression detected — heap or time increased by more than 5%.

✅ Low variance (time: 2.5%, heap: 1.1%)

_{Commit: 71c1377 · Baseline: built from main · Runner: self-hosted bench}

[ColoCarletti]

/bench

[ColoCarletti]

/bench

[ColoCarletti]

/bench

[ColoCarletti]

/bench

[gabrielbosio]

/bench

[gabrielbosio]

/bench

[gabrielbosio]

/bench

[gabrielbosio]

superseded by #466