execution/commitment: commitment metrics as counters, round duration and branch IO bytes

[awskii]

Stacked on #23588 — review that one first. With the metrics reconnected, this replaces the shape they report in: counters and a histogram instead of gauges holding a rate that was already divided by the log interval. A pre-divided gauge fixes the averaging window at the writer, cannot be re-aggregated across instances, and turns a missed scrape into a lost interval rather than a lost sample. rate() belongs in the query.

Changes

• Retired the 13 commit_* gauges.
• Counters: commitment_blocks_total, commitment_txns_total, commitment_keys_total, commitment_folds_total, commitment_unfolds_total, commitment_branch_writes_total, commitment_branch_read_bytes_total, commitment_branch_write_bytes_total; with a kind label, commitment_key_traversals_total{address|storage}, commitment_reads_total{account|storage|branch}, commitment_cache_hits_total and commitment_cache_misses_total.
• commitment_round_duration_seconds histogram plus commitment_rounds_total, observed once per Process on both engines. Buckets run 1ms to 60s, spanning a fast incremental block through a whale fold.
• Branch IO is now measured in bytes, not just operations: reads at unfoldBranchNode, writes at both BranchEncoder sites and at the deferred apply. DeferredBranchUpdate.encoded is filled by the merge inside ApplyDeferredBranchUpdates, so the deferred bytes are summed after it returns, not before.
• Progress.prevCommitment* were read to compute deltas but never rolled forward, so every emission re-sent the cumulative total as if it were the interval's increment. They advance now.

Notes

commitment_keys_total counts distinct keys handed to the trie (updates.Size()); commitment_key_traversals_total counts cell visits. They are different numbers on the parallel engine, which re-walks subtrees on mount+replay: 320 traversals against 80 distinct keys on a 16-account fixture. One counter for both would not be comparable between the two engines, which is why the split exists.

RoundKeys is deliberately not merged between tries — the engine that ran the round owns that number, while a mount worker only sees its own subtree. Merging it multiplied the count by the worker fan-out, giving 320 at four workers for the same 80 keys.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.