blockifier,apollo_batcher: trim OS initial reads to the accessed-key set

[yoavGrs]

Drop initial-reads entries whose keys are not in AccessedKeys (e.g. storage cells read only by a reverted transaction). The OS replays from the read values of the keys it accesses and needs nothing beyond that set, so the extra entries would only bloat the cende blob. Validated by the OS flow tests (113 passing), including all reverted-tx cases.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com

[reviewable-StarkWare]

This change is 

[yoavGrs]

Warning

This pull request is not mergeable via GitHub because a downstack PR is open. Once all requirements are satisfied, merge this PR as a stack on Graphite.
Learn more

• blockifier,apollo_batcher: trim OS initial reads to the accessed-key set #14595 👈 (View in Graphite)
• starknet_os_flow_tests: build OS initial reads via get_os_initial_reads #14594
• apollo_consensus_orchestrator: forward OS initial reads into the cende blob #14593
• blockifier,apollo_batcher: produce OS initial reads behind os_input #14592
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[github-actions]

Artifacts upload workflows:

• Started at 2026-06-22T11:56:41Z
• Started at 2026-06-22T14:02:57Z
• Started at 2026-06-23T07:21:37Z
• Started at 2026-06-23T08:56:15Z
• Started at 2026-06-23T10:41:41Z
• Started at 2026-06-23T13:31:00Z
• Started at 2026-06-23T13:31:01Z
• Started at 2026-06-23T14:42:46Z
• Started at 2026-06-23T18:32:37Z

[cursor]

PR Summary

Medium Risk
Wrong trimming could break OS block replay; scope is limited to os_input paths and is covered by OS flow tests including reverted transactions.

Overview
Shrinks OS initial_reads so only pre-block values for keys in AccessedKeys are kept—dropping reads that never matter for OS replay (e.g. storage touched only by reverted transactions).

Adds StateMaps::trim_to_accessed_keys in blockifier, filters storage, nonces, class hashes, and compiled class hashes against the accessed-key sets. With os_input, the batcher runs this in decision_reached after write_block_accessed_keys and passes the trimmed map in CentralObjects instead of the full execution cache. OS flow tests apply the same trim before building block input so tests match production.

A small comment in apollo_committer documents the os_input vs non-os_input committer request surface; no behavior change there.

^{Reviewed by Cursor Bugbot for commit d865f09. Bugbot is set up for automated code reviews on this repo. Configure here.}

[itamar-starkware]

Suggestion from Claude Code:

Could we add direct unit tests in cached_state_test.rs?

• trim_to_accessed_keys: build a StateMaps with entries across all four maps plus extras absent from accessed_keys; assert the extras are dropped and the hits retained. Cover empty accessed_keys → all empty, and pin the declared_contracts pass-through (it's not trimmed — intentional only because get_os_initial_reads pre-clears it; worth asserting so a future pub caller doesn't get surprised). The test_state_maps helper (~line 700) has the construction pattern.
• get_os_initial_reads: assert the three back-fills (class-hash + nonce per accessed contract, compiled-class-hash per declared class) land, and that declared_contracts comes back cleared.