test: flag e2e_slashing_attested_invalid_proposal as flake under pipelining

[spalladino]

Summary

Add .test_patterns.yml entry to flag e2e_slashing_attested_invalid_proposal > slashes a lazy attester for an invalid checkpoint and clears it on delayed equivocation as a flake. Hit on http://ci.aztec-labs.com/6db6103599cb22e6.

This is a real race under pipelining, not a one-off — same dynamics as the e2e_ha_full fix in spl/fix-ha-full-equivocation-check. Source-level fix is the proper resolution; flake flag is the quick unblock.

Investigation

Failure: TimeoutError: Timeout awaiting honest validator slash offenses for invalid proposal attestation at attested_invalid_proposal.test.ts:404. The honest validator never records an ATTESTED_TO_INVALID_CHECKPOINT_PROPOSAL offense against the lazy validator, so the retry-until times out.

Timeline from the CI log:

t	event
11:38:45.871	validator-1 (bad proposer) broadcasts checkpoint proposal for slot 13
11:38:45.993	bad proposer broadcasts its self-attestation
11:38:46.019	validator-2 (lazy) receives self-attestation (24 ms)
11:38:46.021	validator-2 receives a block proposal — but parent block not found
11:39:09.804	validator-2 finally receives the checkpoint proposal — 23 s after broadcast
11:39:09.808	validator-2 attests (lazy node has skipCheckpointProposalValidation: true, so it signs blindly)
11:39:09.828	validator-1 / validator-3 reject the attestation: Checkpoint attestation slot 13 is not current (14) or next (14) slot (HighToleranceError), peer scored down

Root cause:

1. The p2p CheckpointAttestationValidator (yarn-project/p2p/src/msg_validators/attestation_validator/attestation_validator.ts:56-69) accepts attestations only for targetSlot, nextSlot, or within attestationWindowIntoTargetSlot (~1.5 s) of the current wallclock slot. After that it returns reject with HighToleranceError.
2. The slasher's ATTESTED_TO_INVALID_CHECKPOINT_PROPOSAL detector (yarn-project/validator-client/src/validator.ts:815 handleCheckpointAttestation) is wired only to the gossip path via registerCheckpointAttestationCallback (validator.ts:428). Gossip-rejected attestations never reach it.
3. The lazy validator's attestation was 23 s late because validator-2 received the self-attestation from the bad proposer in 24 ms but didn't see the checkpoint proposal until 11:39:09. Most likely an interaction between block-3 arriving before block-2's processing finishes (parent_block_not_found warnings precede the proposal arrival), with the checkpoint dispatch held behind block validation in the libp2p service.

This is exactly the dynamic described by the commit message on spl/fix-ha-full-equivocation-check (#035ac6a8) — the p2p attestation pool is not a durable record of past-slot attestations under pipelining. The attested_invalid_proposal test sits one step downstream: instead of reading the pool, it relies on the slasher recording the gossip-validated attestation. Same race, same outcome.

Follow-up

Proper source fix: give the slasher's invalid-checkpoint-attestation detector a path that survives p2p gossip rejection — e.g. observe checkpoint attestations matching a slot already flagged for an invalid block proposal before the validator's reject return, scoped to that case. Equivalent in spirit to the HA test rewrite, but on the producing side.

Test plan

• Flake regex matches the observed TimeoutError: Timeout awaiting honest validator slash offenses for invalid proposal attestation from the failure
• CI runs e2e_slashing/attested_invalid_proposal per existing pattern; failures now alert palla on #aztec3-ci instead of failing the build