forkchoice: skip STARK aggregate verify on proofs the filler just produced

[tcoratger]

Summary

Aggregated STARK proofs built by the fixture filler were re-verified by the fork-choice Store on every block and every gossip aggregated attestation — pure duplicated work, since client teams rerun verification when they consume the fixtures.

Thread an opt-in trust flag from the filler down to the two STARK-verify call sites so fresh, filler-produced proofs skip verification while leaving every other code path (client consumers, unit tests, the verify_signatures fixture format) untouched.

• SignedBlock.verify_signatures(..., skip_aggregate_verify=False) — new kw-only flag. Skips the per-attestation STARK aggregate verify. Proposer XMSS verify and structural checks still run.
• Store.on_block(..., trust_proofs=False) — forwards to verify_signatures as skip_aggregate_verify.
• Store.on_gossip_aggregated_attestation(..., trust_proof=False) — gates only the proof.verify(...) call. validate_attestation and validator-index bounds checks still run.
• Fork-choice filler (packages/testing/src/consensus_testing/test_fixtures/fork_choice.py) mirrors the existing valid_signature flag on each spec: valid_signature=True → trust, valid_signature=False → still verify (so explicit invalid-proof tests keep rejecting).

Measured impact

Single-process test-scheme fill of tests/consensus/devnet/:

metric	before	after
wall time	113.59 s	77.10 s (-32%)
verify_aggregated_signatures calls	611	1
verify_aggregated_signatures total	28.5 s	~0 s

The one remaining verify call is the explicit-invalid test that needs the verifier to run. Prod mode (larger STARK params) benefits proportionally.

What does not change

• AggregatedSignatureProof.verify itself — stays pure; the caller decides whether to call it.
• fill.py CLI, filler.py pytest plugin, State.state_transition, State.build_block.
• The verify_signatures_test fixture — calls verify_signatures(...) with the default False, so that fixture format (which exists precisely to exercise the verifier) still runs the full STARK verify.
• Any test under tests/consensus/devnet/. No test was edited.
• Fixture JSON layout and field contents (STARK proof bytes already differ run-to-run because the prover is non-deterministic).

Test plan

• uvx tox -e all-checks clean (ruff, ty, codespell, mdformat).
• uvx tox -e pytest — 3251 passed, 91.74% coverage.
• End-to-end fill on tests/consensus/devnet/fc/ + tests/consensus/devnet/verify_signatures/ — 87/87 passed. Explicit-invalid specs (tests/consensus/devnet/fc/test_gossip_attestation_validation.py, tests/consensus/devnet/verify_signatures/test_invalid_signatures.py) still reject as expected.
• Profiling rerun confirms the expected drop in verify_aggregated_signatures count.
• CI prod-vectors job runtime after merge (expected to drop proportionally from the ~50 min baseline).

New unit coverage

Added 4 minimal tests, folded into existing test files rather than new standalone modules:

• tests/lean_spec/subspecs/containers/block/test_verify_signatures_skip.py (new file, only because containers/block/ had no existing unit tests yet) — 2 tests: skip_aggregate_verify=True bypasses the STARK verifier; proposer-signature verify still runs when the flag is True.
• tests/lean_spec/subspecs/forkchoice/test_store_attestations.py (existing) — 2 new tests: on_block(..., trust_proofs=True) and on_gossip_aggregated_attestation(..., trust_proof=True) both skip the STARK verifier.

🤖 Generated with Claude Code