test(e2e): split node-killing HA test into its own composed suite file

[spalladino]

Round-2 e2e consolidation, PR 9/9 — owns composed/**.

Clock-skew suite stays on the composed HA Postgres cluster

The four Clock Skew and Timezone Safety its in composed/ha/e2e_ha_full.parallel.test.ts exercise PostgresSlashingProtectionDatabase directly against the running 5-node HA cluster's real dockerized PostgreSQL slashing-protection DB:

• TZ-independent duty timestamps (absolute-time storage)
• cleanupOldDuties keeps recent duties when the node clock is 2h ahead
• cleanupOldDuties deletes old duties by DB time even when the node clock is 1h behind
• cleanupOwnStuckDuties keeps recent stuck duties when the node clock is 3h ahead

The property under test is that these DB-clock semantics (CURRENT_TIMESTAMP, timestamptz) are immune to node clock/timezone skew — the node and its database must be able to genuinely diverge in clock and timezone environment for the assertion to mean anything. That only holds against a real, separate Postgres process. An earlier revision moved these into an in-process PGlite file; since PGlite runs inside the Node process, the DB and node can never actually diverge, so that variant is dropped. The tests keep riding the composed cluster's docker Postgres via the shared HaFullTestContext (t.mainPool / t.dateProvider), exactly as before; dateProvider simulates the skewed node clock while the DB uses its own clock. The @electric-sql/pglite dependency added for the PGlite variant is removed from end-to-end (other workspaces that use pglite for unit tests are unaffected).

Order-dependent HA test → own file

The should distribute work across multiple HA nodes test was annotated must run last because it permanently kills every node. Extracted verbatim into composed/ha/e2e_ha_distribute_work.test.ts, which shares the cluster setup with the remaining suite via a new non-test ha_full_setup.ts module (the ~380-line beforeAll/afterAll/helpers moved verbatim into an HaFullTestContext class). Result:

• The killer test gets its own cluster; the ordering contract is gone (no hidden reliance on definition order).
• It has a single it, so it is a plain .test.ts (not .parallel) and runs as one whole-file CI container.
• The remaining e2e_ha_full.parallel.test.ts keeps its 3 non-destructive its (block production, governance voting, keystore reload) plus the clock-skew describe, all order-independent.
• Zero CI runtime change: setup was moved, not rewritten, so behavior is unchanged. No assertions dropped.

Added a .test_patterns.yml flaky entry for the new file mirroring the existing e2e_ha_full one (owner unchanged).

Stranded composed tests — left in place, still excluded

composed/e2e_persistence.test.ts and composed/integration_proof_verification.test.ts are excluded from every CI list and have run nowhere since April 2025. I moved each into a category dir, ran it, and reverted the move because both are broken on the current branch — root causes outside this diff's editable scope:

File	Disposition	Why
integration_proof_verification.test.ts	left in composed/, excluded, documented	Committed fixtures/dumps/epoch_proof_result.json is stale (last regenerated Feb 2026; circuits/VK changed since). bb and the on-chain HonkVerifier both reject the proof (Failed to verify RootRollupArtifact proof!). Needs the fixture regenerated; better relocated to the bb-prover circuit tests.
e2e_persistence.test.ts	left in composed/, excluded, documented	beforeAll no longer completes: the single-node sequencer stalls in checkpoint proposal (waitForAttestationsAndEnqueueSubmissionAsync) and the 600s hook times out. The root cause is in shared setup/sequencer, not this file.

Each file's header comment and the bootstrap.sh exclusion now document the real reason (previously "excluded for unknown reasons"). Decision items for follow-up: regenerate the epoch-proof fixture and relocate integration_proof_verification to bb-prover; fix the single-node checkpoint stall and refile e2e_persistence under single-node/.

Verified locally

• yarn build (full), yarn format end-to-end, yarn lint end-to-end — all clean.
• Confirmed both stranded tests fail on this branch (evidence above), which is why they stay excluded.
• The compose/ha suites need docker (Postgres/Web3Signer) and were not run locally; the e2e_ha_full / e2e_ha_distribute_work split and the clock-skew describe rely on CI. Setup is moved verbatim and each .parallel it is already isolated per container, so runtime behavior is unchanged.