starknet_patricia,starknet_patricia_storage: cut redundant hashmap work in per-block hot paths

[gkaempfer]

Summary

Follow-up performance review of #14850 (starknet_patricia_storage: remove redundant copy in create_db_key), scoped to two more hot-path hashmap inefficiencies found while auditing the surrounding storage/commit code.

1. RocksDbStorage::mset — redundant hashmap re-lookup per key

// before
for key in key_to_value.keys() {
    batch.put(&key.0, &key_to_value[key].0);
}
// after
for (key, value) in key_to_value.iter() {
    batch.put(&key.0, &value.0);
}

mset iterated .keys() and then re-looked-up each value via key_to_value[key] (the Index impl), hashing and probing the map a second time for every one of the O(N_nodes) entries written per block. .iter() yields (key, value) directly — same pattern multi_set_and_delete right below it already uses.

2. FilledTreeImpl::initialize_filled_tree_output_map_with_placeholders — lost capacity hint after the OnceLock refactor

This function used to build its output map with an explicit with_capacity hint (added in #14641 to avoid O(log N) rehash-on-grow). It has since been rewritten to updated_skeleton.get_nodes().filter_map(...).collect() as part of switching placeholders from Mutex<Option<T>> to OnceLock<T>. filter_map's size_hint() lower bound is always 0 (filtering may drop every element), so collect/extend can no longer preallocate, silently reintroducing the rehashing cost the earlier fix removed.

// after
let nodes = updated_skeleton.get_nodes();
let mut filled_tree_output_map = HashMap::with_capacity(nodes.size_hint().0);
filled_tree_output_map.extend(nodes.filter_map(|(index, node)| match node {
    UpdatedSkeletonNode::UnmodifiedSubTree(_) => None,
    _ => Some((*index, OnceLock::new())),
}));
filled_tree_output_map

get_nodes() is backed by a plain HashMap::iter(), so its size_hint().0 is exact — reserving on the unfiltered count is a safe (tight) upper bound for the filtered map, with identical resulting contents.

Test plan

• cargo build -p starknet_patricia -p starknet_patricia_storage
• SEED=0 cargo test -p starknet_patricia — 107 passed
• SEED=0 cargo test -p starknet_patricia_storage — 8 passed, including storage_test::test_storage_concurrent_access::case_1_rocksdb_storage which exercises the mset path
• Reviewed by an independent Opus pass: confirmed filter_map's size_hint semantics, confirmed get_nodes()'s size_hint is exact, confirmed no double-consume/behavior change, no blocking or suggestion findings

🤖 Generated with Claude Code

---

Generated by Claude Code

[cursor]

PR Summary

Low Risk
Behavior-preserving micro-optimizations on iteration and map capacity; no API or semantics changes.

Overview
Per-block Patricia and RocksDB paths shed redundant HashMap work without changing behavior.

RocksDbStorage::mset now walks key_to_value.iter() when building the write batch, so each entry is hashed once instead of iterating keys and re-looking up values via key_to_value[key] (aligned with multi_set_and_delete below it).

initialize_filled_tree_output_map_with_placeholders again preallocates the filled-tree placeholder map: it reserves using the exact size from get_nodes() (a HashMap iterator), then extends the same filter_map logic. A plain collect() after filter_map could not preallocate because the adapter’s size hint lower bound is always 0.

^{Reviewed by Cursor Bugbot for commit 4908dc1. Bugbot is set up for automated code reviews on this repo. Configure here.}

[reviewable-StarkWare]

This change is