Summary

Cherry-pick / port of #21310 to main, combined with the finalized ancestor-descent check that was already on main (from #18956, commit f3ad70f1f7).

main already has a partial spec-faithful rewrite of getFilterBlockTree — the finalized side now does ancestor descent (correct per spec) — but the justified side still deviates from filter_block_tree / get_voting_source in two ways. This PR fixes those.

Spec reference

filter_block_tree:

voting_source = get_voting_source(store, block_root)
correct_justified = (
    store.justified_checkpoint.epoch == GENESIS_EPOCH
    or voting_source.epoch == store.justified_checkpoint.epoch
    or voting_source.epoch + 2 >= current_epoch
)
finalized_checkpoint_block = get_checkpoint_block(store, block_root, store.finalized_checkpoint.epoch)
correct_finalized = (
    store.finalized_checkpoint.epoch == GENESIS_EPOCH
    or store.finalized_checkpoint.root == finalized_checkpoint_block
)

get_voting_source:

if current_epoch > block_epoch:
    return store.unrealized_justifications[block_root]   # pull-up
else:
    return head_state.current_justified_checkpoint        # realized

Changes

1. Voting source selection: was unrealized-with-fallback unconditional. Now branched on current_epoch > block_epoch as the spec requires — unrealized only for prior-epoch blocks; current-epoch (or future) blocks use the block's realized current_justified_checkpoint.

2. correct_justified: was voting_source.Epoch >= justifiedCheckpoint.Epoch with an is_previous_epoch_justified-gated +2 fallback. Now the three flat disjuncts straight from the spec: epoch == GENESIS or voting_source.epoch == justified.epoch or voting_source.epoch + 2 >= current_epoch. (The previous >= accepts cases spec rejects; the previous gated +2 rejects cases spec accepts.)

3. correct_finalized: unchanged — keeps the ancestor-descent check from f3ad70f1f7 (finalizedCheckpoint.Root == f.Ancestor(blockRoot, finalizedSlot).Root).

Background

Original observed symptom on bloatnet was ~30-block execution unwinds every ~6 minutes (every epoch boundary). Root cause: getFilterBlockTree rejected the leaf at each new epoch's first slot, Caplin's head fell back to an older block, and execution unwound to follow. Both pre-existing deviations contributed:

• Unconditional unrealized fed a stale checkpoint into the check for current-epoch leaves.
• Strict Equal (or >= with the gated fallback) rejected leaves whose voting source's epoch was within 2 of current_epoch but not equal to justified.epoch.

Issue: #21301
Original performance-branch PR: #21310

Test plan

• Spec-faithful: code matches the consensus-specs Python source line-for-line
• go build ./cl/phase1/forkchoice/ clean
• Bloatnet runtime validation on performance (cl/phase1/forkchoice: align getFilterBlockTree with consensus spec #21310): 1 hour at chain tip, 5+ epoch boundaries, zero unwinds, zero filterTree rejects, Caplin lagSlots=0 on all 172 FCU events
• CI green on main
• Caplin unit tests pass