[Acentelles]

Summary

This PR adds two LLVM options to the guest-codegen build defaults of the recursion verifier guest:

• -Cllvm-args=-split-spill-mode=size: the register allocator splits live ranges around spills in the mode that minimises code size.
• -Cllvm-args=-riscv-disable-cost-per-use=true: the RISC-V register allocator stops weighting registers by their encoding cost.

Both go into the same default string as the existing inline threshold, GVN hoisting, DFA jump threading, ExtTSP block placement and no-tail-merging flags, still with one codegen unit. A caller value for ZEROOS_GUEST_RUSTFLAGS still wins, and RECURSION_GUEST_CODEGEN=off restores the stock guest build. This is a host-side source change only: the inner guests, the Jolt proof format and the Akita proof are unchanged.

This PR stacks on #2143. Please review only the last two commits. GitHub also shows #2143's commits until #2143 merges.

Measured effect

These are total trace rows (real plus virtual) of the recursion example's verifier guest, which verifies one Akita proof of an inner Jolt proof.

Both changes together, on this public tree. "Without" is #2143's published head 85bee6f27. "With" is this branch, built from a fresh clone with Akita fetched by Cargo from the pin.

Inner proof	Without (#2143)	With (this PR)	Difference
$2^{12}$	58,017,123	57,813,912	-203,211
$2^{20}$	148,001,352	147,659,408	-341,944

• The "without" figures are perf(recursion): cut the recursion verifier guest to 44.9M trace rows at inner 2^12 (43.2% below #1911) #2143's published figures.
• The "with" figures were measured on a Ryzen 9 5950X host, together with:
  • a rejected tampered proof (48,790,143 rows);
  • a repeat with a regenerated proof (same rows and ELF);
  • Clippy with -D warnings in the literal, default and no-default-features graphs.
• The proofs are byte-identical to perf(recursion): cut the recursion verifier guest to 44.9M trace rows at inner 2^12 (43.2% below #1911) #2143's, and the Akita proof stays at 72,672 bytes ($2^{12}$) and 75,817 bytes ($2^{20}$).
• 57,813,912 rows is 26.85% below feat(recursion): Akita field-inline verifier guest on main #1911's 79,032,888.

Each change alone. These were measured on the equivalent build that uses a local Akita checkout. That build runs a constant 14 rows above the public tree at both sizes, with the same proofs.

Change	Inner proof	Without	With	Difference	Hosts
-split-spill-mode=size	$2^{12}$	58,017,137	57,927,886	-89,251	Ryzen 9 5900XT, Ryzen 9 5950X
	$2^{20}$	148,001,366	147,794,518	-206,848	same
-riscv-disable-cost-per-use=true (on top of the first)	$2^{12}$	57,927,886	57,813,926	-113,960	Ryzen 9 5950X (two hosts), EPYC 9124, Ryzen 9 7950X
	$2^{20}$	147,794,518	147,659,422	-135,096	same

The Akita proof size is unchanged for each change.

How to reproduce

git clone -b recursion-guest-codegen-spill-costperuse https://github.com/Acentelles/jolt
cd jolt
cargo build --locked --profile ci -p recursion --features akita,field-inline,ntt-inline
target/ci/recursion generate --example fibonacci --workdir wd
RAYON_NUM_THREADS=1 target/ci/recursion trace --embed --expect accept --example fibonacci --workdir wd

This is #2143's command with this branch. Inner $2^{12}$ is the example's default input. The $2^{20}$ figures used the same host-only input override as #2143 (RECURSION_FIB_N=74858), which is not in this branch.

[Acentelles]

Folded into #2143.