bpf-bench: integrate VerifyCompressedProof CU measurements (s116 follow-up)

[kh0ra]

Why

We have host-side criterion benches for alt_bn128 compression (sessions 112 / 114), but no SBF measurement of the actual on-chain VerifyCompressedProof (instruction tag 0x03, session 116). Without this we can't claim a concrete CU number for the compressed-path.

Sub-tasks

• Add bench targets in mosaic-bench::bpf_bench::TARGETS[] for each compressed verifier (Groth16, PLONK, Halo2, HyperPlonk, Nova)
• Baseline + hard cap per system
• Document in docs/compute-unit-budget.md
• Update README CU table to show "canonical vs compressed" columns

Note

Once measured we can claim "decompression adds X CU on top of baseline" with a real number, not an estimate.

[kh0ra]

Partial close shipped at 40b6967 (94cedd5..40b6967).

5 new TARGETS landed in crates/mosaic-bench/src/bin/bpf_bench.rs:

• groth16_compressed_mul_circuit_1pi — hard cap 380K
• plonk_compressed_mul_circuit_1pi — hard cap 1.3M
• hyperplonk_kzg_compressed_scaffold — hard cap 860K
• halo2_kzg_compressed_scaffold — hard cap 960K
• nova_folding_compressed_scaffold — hard cap 1.35M

TARGETS count 7 → 12.

Still open before this issue can close:

1. Pinned baselines — every new entry ships with baseline_cu = 0. First successful run with the SBF artifact in target/deploy/ records the real number; that PR updates the TARGETS entries.
2. README CU table — currently shows canonical-only columns. Needs a 'canonical / compressed' split column once baselines land.
3. docs/compute-unit-budget.md — needs entries for the new compressed-path systems.

Same commit also ships the SBF integration test coverage (4 new tests, count 13 → 17) that exercises the same dispatch arm — covered separately at 94cedd5.

[kh0ra]

Measurement half done at fd5f97b — all 11 dispatch arms have real CU baselines

First sweep where every arm was measured end-to-end on the solana-program-test VM (unblocked by the borsh-1.5.7 fix in #88). Build: cargo build-sbf --tools-version v1.52.

Target	Measured CU	Cap	Status
groth16_bn254_mul_circuit_1pi	84,027	180K	✅ (+0.54% vs prior)
groth16_batch_n5	259,772	300K	✅ (+0.53%)
plonk_bn254_mul_circuit_1pi	973,388	1.1M	✅ (+0.51%)
hyperplonk_kzg_scaffold	900,750	1.05M	✅ first measurement
halo2_kzg_scaffold	824,074	950K	✅ first
nova_folding_scaffold	289,899	360K	✅ first
groth16_compressed	146,620	190K	✅ first
plonk_compressed	1,005,100	1.2M	✅ first
hyperplonk_compressed	928,039	1.1M	✅ first
halo2_compressed	857,503	960K	✅ first
nova_compressed	316,580	400K	✅ first
fri_stark_scaffold	—	7.8M	⚠ fixture gap → #76

Production verifiers re-measured within 0.6% of their pinned baselines — arithmetic is stable.

3 estimate-derived caps corrected against real data: HyperPlonk (660K→1.05M) and Halo2 (760K→950K) were under-capped (host estimate under-counted 78%/42%); Nova was over-capped 4× (1.15M→360K) so the bench now actually catches regressions.

bpf-bench made resilient: records per-target failures and continues instead of aborting on the first error, so one over-budget scaffold can't hide the production baselines.

fri_stark large-shape scaffold fails on-chain verification with Custom(0x2F) — build_stark_scaffold_fixture() doesn't construct valid Merkle paths at that shape. The depth-zero STARK shape in verify_proof_sbf passes, so the dispatch + verifier are sound; this is a bench-fixture-builder gap, tracked with #76.

Full table + estimate-vs-measured analysis in docs/compute-unit-budget.md. mosaic-soak pinned baselines synced to match.

Remaining for #84: wire the bench into CI (with the v1.52 toolchain pin from #88) so baselines are enforced on every push, + fix the fri-stark bench fixture (#76).