refactor(e2e): fast block build config in e2e tests

[spalladino]

Motivation

AnvilTestWatcher polls every 200ms and silently mutates global state from three different RunningPromises — proving, dateProvider sync, and L1 time warping. The wiring difference between e2e and sandbox makes failures hard to diagnose. Concrete examples this work was motivated by:

• A two-dateProvider-read race in Sequencer.work where targetSlot could compute to slot + 2 because syncDateProviderToL1IfBehind warped time between the two reads.
• e2e_pruned_blocks regressions where finalized stalled at 0 because markAsProven and the archiver's finalized tag advanced out of order.
• Blacklist huge-time-warp wedges in the pipelined publisher: the in-flight checkpoint references pre-warp state while the watcher keeps operating on post-warp state.

Deleting the watcher from e2e in favor of deterministic per-responsibility replacements gets us:

• L1 time advances on a known cadence (anvil_setIntervalMining), no surprise warps.
• Proven tip moves only when a real prover or EpochTestSettler advances it; no async loop racing the archiver.
• dateProvider mutates only when a cheat code explicitly warps L1, atomically with the RPC.

The AnvilTestWatcher class stays in the tree so the sandbox / local-network entrypoint can keep using it. This PR only changes the e2e fixture wiring.

What's landed

Foundation

• fix(ethereum): atomically sync TestDateProvider in cheat-code mine/warp paths — mine(), evmMine(), mineEmptyBlock() sync TestDateProvider after the RPC. Removes redundant manual setTime at three e2e call sites. setNextBlockTimestamp deliberately remains un-synced (it schedules but doesn't mine; sync happens on the next mine).
• feat(e2e): add FAST_E2E_SETUP_OPTS preset — single-block fast-mode timing (3.75s execution budget per slot, 20s pipelined publishing deadline, 48s epoch). Plus a stdlib/timetable/index.test.ts unit test that verifies the derived numbers.
• refactor(e2e): unconditional anvil interval mining + drop AnvilTestWatcher — fixtures/setup.ts integration: drop automineL1Setup, switch to unconditional setIntervalMining(ethereumSlotDuration) for anvil after L1 deploy, default-skip watcher.start() for anvil chains (disableAnvilTestWatcher ?? isAnvilChain). Removes redundant setIsMarkingAsProven(false) calls from 8 sites.
• refactor(e2e): use EpochTestSettler in place of CheckpointAutoProver — reuses the existing sandbox-path orchestrator (yarn-project/aztec/src/local-network/local-network.ts:213) instead of the bespoke per-checkpoint helper the PR originally added. Advances both the outbox hash and the proven tip once per completed epoch.
• test(e2e): add e2e_fast_config smoke test — canary exercising three invariants: (a) chain advances under interval mining + pipelining, (b) 20 sequential dependent txs land in distinct blocks (single-block mode keeps batching off), (c) proven tip advances within two epochs (EpochTestSettler works).

Migrations (49 tests)

• 10 first-batch single-sequencer tests (golden-path, no special timing semantics).
• 22 second-batch single-sequencer tests.
• e2e_amm (separate commit because it had a third setup() arg the batch-2 agent missed).
• 14 tests reachable through TokenContractTest / NestedContractTest, which gained a opts?: Partial<SetupOptions> parameter (caller opts first, fixture invariants like metricsPort / fundSponsoredFPC / skipAccountDeployment override; mirrors the existing FeesTest.setup(opts) pattern).
• e2e_2_pxes and e2e_lending_contract (initially deferred per the design plan, then re-verified safe).

Held back from this PR

• e2e_cheat_codes — exercises setNextBlockTimestamp and mine() directly. The cheat-code audit deliberately left setNextBlockTimestamp un-synced; this test reads dateProvider.now() between schedule and mine, so it needs a hand-audit before migration.
• BlacklistTokenContractTest (7 suites) — huge time warps wedge the pipelined publisher, and they also hit a separate L1ToL2MessagesNotReadyError bug (see "Out of scope" below). Blocked on those source fixes.

Explicitly out of scope (separate PRs)

These came up during the work and are tracked but not addressed here:

• canProposeAt simulation regression under pipelining. The simulation drops lastArchiveRoot and the pending-tip slot override; a fix exists on origin/sp/publisher-simulation-cleanup and needs forward-port to merge-train/spartan. Blocks e2e_contract_updates, composed/web3signer/e2e_multi_validator_node_key_store, and HA governance opt-in.
• AztecNodeService.simulatePublicCalls queries unsealed checkpoint. With inboxLag = 2, the simulator calls getL1ToL2Messages(proposedCheckpoint + 1) on a checkpoint that hasn't been sealed yet, raising L1ToL2MessagesNotReadyError. Pointer: yarn-project/aztec-node/src/aztec-node/server.ts:1519-1524. Preferred fix: a positive messagesSealed(checkpoint) check on the message store, gating the next-checkpoint-message-append on it. Blocks 7 blacklist suites + e2e_bot + e2e_avm_simulator + e2e_fees/{private_payments,failures}.
• Publisher funder loop stops firing after cycle 1. RunningPromise cycle in yarn-project/ethereum/src/publisher_manager.ts:84-89. Hypothesis: shared viem client between an in-flight monitorTransaction and cycle-2's getSenderBalance hangs the loop. Blocks e2e_publisher_funding_multi.
• Sequencer.invalidateInFlightCheckpointsBefore(slot) — needed if any test wants to do huge L1 warps that would orphan an in-flight checkpoint. Currently the in-flight job is silently kept and times out.

Status

Audited end-to-end by an LLM review pass + a /codex round-2 review on the final state of the branch. No CI-blocking issues flagged. Mark non-draft to trigger e2e CI. Consider ci-no-fail-fast for the first run so all failures surface at once.