perf(programs): build the ethrex guest with thin LTO

[MauroToscano]

What

Add a [profile.release] with lto = "thin" to the ethrex guest crate
(executor/programs/rust/ethrex/Cargo.toml). The crate previously had no
[profile.release] at all, so it built with cargo's release defaults
(lto = false, codegen-units = 16).

Why

Thin LTO measurably lowers the number of instructions the VM executes for a
real ethrex block, at no runtime cost — execution in this VM is priced per
executed instruction, so the small ELF-size increase is free.

Measured on the committed fixtures (same toolchain, same inputs):

fixture	baseline (cycles)	thin-LTO (cycles)	Δ cycles	Δ %	Keccak	Ecsm
ethrex_bench_4.bin	3,841,262	3,750,291	−90,971	−2.37%	174	16
ethrex_bench_16.bin	11,616,950	11,352,127	−264,823	−2.28%	356	64

The syscall invariants are identical before and after (Keccak 174 / Ecsm 16 for
_4, Keccak 356 / Ecsm 64 for _16) — the guest does exactly the same work;
LTO just compiles it into fewer executed instructions.

Methodology

Deterministic A/B on a dedicated 32-core rig, same toolchain
(nightly-2026-02-01, -Z build-std, riscv64im-lambda-vm-elf), cycles read
via cli execute --cycles on the committed fixtures. The measurement noise
floor is exactly 0: the pristine baseline was built twice in separate
CARGO_TARGET_DIRs and produced byte-identical ELFs, so the deltas above are
the whole signal.

Size

Guest ELF grows 3.64 MB → 3.82 MB (+4.9%). This is free here: the VM charges
per executed instruction, not by ELF size.

Follow-up (out of scope)

A tuned variant — thin LTO plus -C llvm-args=-inline-threshold=2500 /
-unroll-threshold=2000 — measured a further ~2× improvement
(ethrex_bench_16 11,035,843, −5.00% total; ethrex_bench_4 3,616,998,
−5.84% total) at +66% ELF size. That inline/unroll tuning is deliberately left
out of this PR and can land as a separate change.

[MauroToscano]

/bench

[github-actions]

Benchmark — ethrex 20 transfers (median of 3)

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

Metric	main	PR	Δ
Peak heap	73028 MB	71902 MB	-1126 MB (-1.5%) ⚪
Prove time	37.771s	37.731s	-0.040s (-0.1%) ⚪

✅ No significant change.

🔬 Looks like a small speedup (-0.1%) — below what 3 runs can confirm. Comment /bench-abba to run the drift-free ABBA tiebreaker (paired-t CI + exact Wilcoxon). Note: it occupies the bench server for ~30–40 min.
Optional pair count: /bench-abba 32 (20 resolves ~1%, 32 for ~0.6%).

✅ Low variance (time: 0.8%, heap: 1.8%)

_{Commit: 9461f88 · Baseline: cached · Runner: self-hosted bench}