Gloas/Heze mainnet sanity/slots/historical_accumulator spec test is ~7.5x slower than pre-Gloas forks

[ensi321]

Summary

sanity/slots/pyspec_tests/historical_accumulator takes ~23-24s for gloas and heze on
the mainnet preset, vs ~1-6s for every pre-gloas fork. The sanity slots timeout is
30000ms (packages/beacon-node/test/spec/presets/sanity.test.ts:49), so these two cases
run at 76-81% of the budget and are a timeout-flake risk on slower CI runners.

Measurements

mainnet preset, sanity/slots/pyspec_tests/historical_accumulator, local run @ v1.7.0-alpha.13:

fork	duration
phase0	1003ms
bellatrix	2241ms
deneb	2598ms
electra	3032ms
fulu	3142ms
altair	5003ms
capella	5847ms
gloas	22948ms
heze	24234ms

gloas is 7.3x fulu; heze is 7.7x fulu. heze is consistently ~1.3s above gloas.

On the minimal preset all forks finish in <=27ms (gloas 26ms, heze 27ms), so this is
mainnet-only.

Why mainnet only

The test runs process_slots(state, state.slot + SLOTS_PER_HISTORICAL_ROOT) — 8192 slot
transitions on mainnet vs 64 on minimal.

Likely cause

Two gloas-specific changes make each of those 8192 slot transitions more expensive:

1. packages/state-transition/src/slot/index.ts:34-40 — every post-gloas slot writes into
   executionPayloadAvailability, a BitVector[SLOTS_PER_HISTORICAL_ROOT] (8192 bits on
   mainnet):

   if (fork >= ForkSeq.gloas) {
     // Unset the next payload availability
     (state as CachedBeaconStateGloas).executionPayloadAvailability.set(
       (state.slot + 1) % SLOTS_PER_HISTORICAL_ROOT,
       false
     );
   }
   That is 8192 ViewDU writes into a large bitvector, each dirtying nodes that must be
   re-merkleized on commit.

2. The gloas BeaconState is a ProgressiveContainer (EIP-7688), which has a different
   merkleization cost profile than the fixed-depth containers used pre-gloas. heze widens
   it further, consistent with heze being slower than gloas.

Not yet profiled — the split between (1) and (2) is unknown.

Impact

• Timeout-flake risk: 76% (gloas) / 81% (heze) of the 30s budget.
• Inflates total mainnet spec-test wall clock.
• Any future fork inherits the regression.

Reproduce

pnpm vitest run --project spec-mainnet
packages/beacon-node/test/spec/presets/sanity.test.ts -t historical_accumulator

Proposed short term

Skip the two cases with a pointer to this issue, rather than raising the timeout —
raising it hides the regression. See #9763 discussion.

Investigation ideas

• Profile to attribute cost between the bitvector write and progressive-container hashing.
• Consider deferring/batching the executionPayloadAvailability reset so process_slots
  does not commit a large bitvector every slot.
• Compare against a fixed-depth BitVector control to isolate the EIP-7688 contribution.