refactor(e2e): consolidate the multi-node and single-node test categories

[spalladino]

Motivation

The e2e suite under end-to-end/src/ had accumulated ad-hoc base classes with overlapping
responsibilities and large multi-describe files that were hard to navigate. The A-1175 survey settled on
consolidating the suite into a few setup/lifecycle categories, each backed by one base class that owns the
environment, with domain behavior composed on top. This PR delivers the single-node and
multi-node categories, organizes each into behavior-named subfolders, and breaks the suite into small,
one-describe files.

Approach

single-node and multi-node are sibling categories. SingleNodeTestContext owns the environment (one
sequencer + optional fake prover, prover lifecycle, epoch/proof/reorg waiters); MultiNodeTestContext extends
it, adding a validator committee, gossip helpers, and validator-registration sugar. Every file with more than
one top-level describe was exploded into a subfolder of one-describe files, with shared describe-level
setup pulled into a co-located setup.ts.

Hierarchy: before → after

Before — one flat directory plus two strays:

e2e_epochs/                 ~30 epochs_*.test.ts  +  epochs_test.ts (base context)
e2e_p2p/duplicate_attestation_slash.test.ts
e2e_p2p/duplicate_proposal_slash.test.ts

After:

single-node/                          # one sequencer (+ optional fake prover)
  single_node_test_context.ts         # base context + shared timing profiles
  proving/                            # epoch/proof lifecycle is the subject
    world_state_pruning  
    empty_blocks  
    long_proving_time  
    multi_proof
    optimistic.parallel  
    proof_fails.parallel  
    cross_chain_public_message  
    upload_failed_proof
  partial-proofs/                     # mid-epoch / multi-root / Outbox semantics
    single_root  
    multi_root
  l1-reorgs/                          # split along its describes
    blocks.parallel  
    messages.parallel
  recovery/                           # reorg + pending-chain recovery
    manual_rollback  
    sync_after_reorg  
    prune_when_cannot_build
  misc/
    missed_l1_slot                    # single-node sequencer sync/timetable regression

multi-node/                           # N validators on mock gossip
  multi_node_test_context.ts          # also owns validator registration (harness folded in)
  block-production/                   # happy-path committee production
    simple  
    high_tps  
    first_slot  
    proof_boundary.parallel  
    proposed_chain.parallel
    cross_chain_messages.parallel  
    deploy_and_call_ordering  
    blob_promotion  
    redistribution.parallel
  recovery/                           # detect a bad/withheld/conflicting proposal → recover
    proposal_failure_recovery.parallel  
    pipeline_prune  
    equivocation_recovery
  invalid-attestations/
    invalidate_block.parallel         # invalid checkpoints detected/invalidated, chain progresses
  high-availability/
    ha_sync  
    ha_checkpoint_handoff
  slashing/                           # pure offense detection
    duplicate_proposal  
    duplicate_attestation

Criteria for the hierarchy

• Top level = topology / setup model: single-node (one sequencer) vs multi-node (validator committee
  • gossip) — it names the environment a test needs, never the feature it happens to touch.
• Second level = the primary system behavior under test — not the shared setup, and not a flag. "Prover"
  is a flag, so a multi-validator test isn't a "proving" test just because a prover participates (hence
  multi-node has no proving/ folder, single-node does).
• Third level only for behavior that is still genuinely exceptional. MBPS and pipelining are now default
  production traits, not toggleable features, so they no longer earn a folder or a filename.
• Folder homogeneity: every test in a subfolder uses that subfolder's context — placement never follows
  helper/context ownership (e.g. prune_when_cannot_build stands up one solo sequencer, so it's a single-node
  test even though it came from the multi-node setup).
• One top-level describe per file; shared setup in a co-located setup.ts; the .parallel suffix iff a
  file has more than one it (preserving CI's per-it job splitting without extra anvils).

Tests removed / replaced

• One test removed (net it count 64 → 63): the MBPS case "builds multiple blocks per slot with
  transactions anchored to checkpointed block", dropped as redundant with the retained "...anchored to
  proposed blocks" variant (now block-production/proposed_chain).
• No other test removed. The two e2e_p2p duplicate-slash tests were moved, not deleted, into
  multi-node/slashing/ (and switched to mock gossip). Several files were merged (combining files while
  keeping every distinct it): mbps/l2_to_l1+l1_to_l2 → cross_chain_messages.parallel;
  mbps/proposed_anchor+non_validator_sync → proposed_chain.parallel;
  prune/missed_l1_publish+orphan_block_prune → proposal_failure_recovery.parallel. l1_reorgs.parallel
  was split along its describes into l1-reorgs/blocks.parallel + messages.parallel.
• A few it titles changed with their files: the two block_building "builds blocks without any errors"
  tests became "builds simple/high-tps blocks..."; "manually rolls back" → "...to an unfinalized block";
  and the pipelining test was renamed blob_promotion ("promotion-disabled node fetches blobs while peers
  skip them, and the checkpoint proves"), trimmed to its unique blob-promotion + proving assertions (the
  redundant MBPS/pipelining-offset re-checks are still asserted by recovery/pipeline_prune).

Helpers introduced

• SingleNodeTestContext — base context (environment, prover lifecycle, epoch/proof/reorg waiters); owns
  the shared timing profiles REORG_TIMING_BASE, FAST_REORG_TIMING, MULTI_VALIDATOR_REORG_TIMING,
  MULTI_VALIDATOR_BLOCK_PRODUCTION_TIMING, and WIDE_SLOT_TIMING (the 72s wide-slot cadence for
  prover-backed multi-block-per-slot tests).
• MultiNodeTestContext — extends the above with the validator set, committee/gossip helpers, and the
  validator-registration API (validatorAt/addressAt/privateKeyAt/createValidatorNodeAt,
  getSlashingContracts(), and a slasherEnabled setup preset).
• fixtures/wait_helpers.ts — intent-revealing waiters: waitForBlockNumber, waitForProvenBlock,
  waitForNodeCheckpoint, waitForNodeProvenCheckpoint, waitForTxs.
• Co-located setup.ts per multi-file subfolder (proving, partial-proofs, l1-reorgs, recovery
  under single-node; block-production, slashing under multi-node). The block-production setup factors a
  shared buildValidatorCluster spine into two presets: setupSimpleBlockProduction (lean, prover-less
  liveness/throughput canary) and setupBlockProductionWithProver (prover-backed, with a wallet + contract +
  fail-event tracking wired up for content/proving assertions).

A follow-up pass pulled the most-repeated hand-rolled steps up into higher-level helpers so test bodies
read as a sequence of named steps (the long tail of smaller helpers is omitted here):

• warpToBuildWindowForSlot / waitForBuildWindowForSlot (on the context) — the pipelining "build
  window" warp/wait (one L1 block before an L2 slot), replacing the getTimestampForSlot(slot) - L1_BLOCK_TIME + warp arithmetic copied across the consensus/recovery/HA tests.
• proveAndSendTxs (test-wallet/utils.ts) — pre-prove a batch of interactions and send them
  NO_WAIT, collapsing the timesAsync(proveInteraction) + Promise.all(send) pair duplicated in ~13 tests.
• startSequencers / watchNodeSequencerEvents (on the context) — start and watch a committee of node
  sequencers without the nodes.map(getSequencer) boilerplate.
• waitForAllNodesToReachCheckpoint / waitForOffenseOnNodes (on MultiNodeTestContext) — the
  multi-node fan-out polls for a checkpointed tip and for slash-offense convergence.
• assertMultipleBlocksPerSlot consolidated onto one context method that owns both wait modes
  (some-checkpoint-has-N-blocks and checkpointed-tip-reaches-block-X), replacing three divergent copies.

Semantic changes (setups that differ)

• Slashing/equivocation tests run on the in-memory mock gossip bus instead of real libp2p (they moved out
  of e2e_p2p) — deterministic and faster.
• prune_when_cannot_build migrated from MultiNodeTestContext to SingleNodeTestContext — it always
  stood up a single solo sequencer (zero extra validators), so it now uses only the base-class environment.
• Hand-rolled retryUntil checkpoint / block-number polls replaced by the intent-revealing wait-helpers
  above. (retryUntil remains the primitive for tx-status and other custom predicates.)
• Incidental per-test timing values converged onto the shared timing profiles owned by
  SingleNodeTestContext.

Test fixes introduced

• Wait-helper falsy-0 fix (bbfbe69): waitForNodeCheckpoint / waitForBlockNumber now wrap their
  matched value so a correct answer of 0 is no longer read as falsy by retryUntil (which otherwise polls to
  timeout). This was the actual cause behind the l1-reorgs proof-reorg cases that were temporarily skipped
  during development (A-1266); they now run.
• proving/optimistic.parallel deflake: the top-tree-proving gate now captures the live prover session
  directly instead of polling for an awaiting-checkpoints job — the session flips to awaiting-root before
  the hook fires, so the old predicate never matched and the gate never engaged, letting proving race the
  prune. Test-only.

API changes

Test-infrastructure only (no product/public API). EpochsTestContext → MultiNodeTestContext (extends the new
SingleNodeTestContext); the standalone ValidatorRegistrationHarness is gone (folded into
MultiNodeTestContext); e2e_epochs/ removed and the epochs_ prefix dropped; all importers, the CI
test-discovery globs in bootstrap.sh, and the .test_patterns.yml path entries were updated to match.

Part of A-1176 (consolidation roadmap) / A-1064.