fix(prover): emit missing marker on continuation verify paths

[Oppen]

Summary

• verify_epoch's per-epoch verify and verify_global (continuation.rs) call Verifier::multi_verify_views directly, without the STEP_AIRS_AND_BUS_BALANCE_DONE marker that prover/lib.rs's monolithic verify_proof_parts emits before its own multi_verify_views call.
• The recursion-block profile test (test_recursion_profile_blowup4_block) buckets cycles by the latest marker observed, so on the continuation path the "multi_verify setup (transcript replay phase A/B, per-table fork)" step reported a flat 0% — its real cost was silently folded into whichever bucket was already active (airs_and_bus_balance for the first epoch, step4:openings carried over for later epochs/the global proof).
• Adds the same marker call at both continuation call sites, matching the monolithic path.

Test plan

• cargo check -p lambda-vm-prover --lib
• make test-profile-recursion-block — step 2 now reports 17,041,080 cycles (1.16%) instead of 0

[MauroToscano]

⚠️ AI-generated review. Produced by an automated agent and posted from this account — not written by hand. Findings below are leads to verify, not confirmed conclusions.

Adds STEP_AIRS_AND_BUS_BALANCE_DONE at the two continuation multi_verify_views call sites so the block profile stops reporting a flat 0% for "multi_verify setup". Placement matches verify_proof_parts (after compute_expected_commit_bus_balance_view, before the call); step_marker is addi x0, x0, N gated on cfg(all(feature = "profile-markers", target_arch = "riscv64")), so there is no host-side or semantic surface — no transcript touch, no new panic path, nothing driven by proof data. seed() and global_transcript(...) are call arguments, so they evaluate after the marker and land in bucket 2 as on the monolithic path. Correct as far as it goes, but the attribution fix is only half complete.

prover/src/continuation.rs:809 — the marker reclaims the [2 -> 3] window, but everything in verify_epoch before it is still billed to the latched bucket, which for every epoch after the first is bucket 6. verifier.rs:610 leaves STEP_VERIFY_TRACE_AND_COMPOSITION_OPENINGS set when multi_verify_views returns, and the epoch loop (continuation.rs:1271-1295) emits nothing between iterations. So for epochs 1..n-1, build_epoch_airs — including register::compute_precomputed_commitment_with_fini, an FFT+Merkle recompute per epoch — plus the first seed() (a full-ELF keccak inside absorb_statement) and the phase-A replay in compute_expected_commit_bus_balance_view are all charged to "6. step4:openings". Only epoch 0's pre-work reaches bucket 1, via the guest's single STEP_DECODE_DONE. The 17.0M cycles reclaimed here are themselves dominated by the second seed() keccak, so the residual is the same order of magnitude: "1. airs_and_bus_balance" under-reports by roughly a factor of n and "step4:openings" is inflated by the same amount — on exactly the numbers used to prioritize verifier work. Cheap follow-up: emit a bucket-1 marker at the top of verify_epoch; a dedicated STEP_EPOCH_SETUP is preferable to reusing STEP_DECODE_DONE, whose name would then be wrong. verify_global:971 has the same shape but is cheap on the profiled path, since the guest supplies page_commitments and global_memory_configs_classify_only skips the per-page genesis recompute.

Nits:

• No test guards the two instrumented paths: run_profile_from (prover/src/tests/recursion_smoke_test.rs:409) asserts the attestation but nothing about buckets, and the marker-order regression test drives only recursion-min (setup_guest_run -> Preset::artifact_stem()), so a future drop reappears as the same silent 0% row — asserting buckets[1..=6] are all nonzero at the end of run_profile_from holds for both guests and is preferable to a continuation-guest variant, since setup_block4_blowup4_guest_run panics outright when the block fixture is absent.
• prover/src/tests/recursion_smoke_test.rs:778 — the documented convention ("1 -> 2 once each, then N repetitions of 3,4,5,6", transitions {1->2, 2->3, 3->4, 4->5, 5->6, 6->3}) is now monolithic-only: on the continuation guest marker 2 fires n+1 times and 6->2 is a legal edge; harmless today, but this comment is the only written record of the convention.