[TomWambsgans]

Faster table sumcheck

Each table's constraint is a quadratic polynomial in its columns, for example P = 3a + 5b + a·c + a·d. Each round, the sumcheck folds the table (halves its rows, mixing pairs with the round's challenge) and evaluates P on the folded rows. It used to fold every column and evaluate P row by row.

Now P is first rewritten with its products grouped by a shared column: P = L + a·R, with L = 3a + 5b and R = c + d. Folding is linear, so folding L gives the same result as folding a and b and then combining them. The sumcheck therefore folds just L, a and R, built once from the original columns when the table joins. On real tables that's a handful of columns instead of dozens, so each round moves far less memory. At the end, each original column is evaluated once to give its claim.

The proof is unchanged, byte for byte. Only the prover changes, and the three verifiers are untouched.

Also in this PR

• Removes the old base-field path (ColVal, eval_constraint_k, RowForm) and the primitives only it used.
• JUMP's identities are now stated as data (Table::constraint) instead of a closure.

Adapted from gattaca-com/leanVM's keccak-opt branch, without their bus rewrite (for now).

Benchmarks

Proving time and peak memory against main, paired A/B. Aggregate is --xmss 900, recursion is --n 2 --xmss-per-leaf 900.

machine	aggregate	recursion	peak memory (aggregate)
M4 Max	−1.6%	−4.0%	6.31 → 5.78 GiB
Zen 4	−6.2%	−5.1%	5.00 → 4.51 GiB
Zen 5	−4.8%	−3.2%	5.07 → 4.60 GiB
Zen 4, AVX2 only	−2.6%	−2.8%	4.92 → 4.51 GiB

🤖 Generated with Claude Code