Optimize batch reconstruction protocol for improved concurrency

[Mikerah]

Summary

• Replace nested Arc<Mutex<HashMap>> with Arc<DashMap> for single-lock session access
• Add HashSet for O(1) deduplication instead of O(n) linear search
• Add AtomicBool guards to prevent race conditions during interpolation
• Move expensive interpolation operations outside lock scope
• Convert clear_entire_store, clear_store, and get_or_create_store to sync methods

Performance Improvements

Operation	Before	After
Session lookup	2 locks	1 lock
Deduplication	O(n)	O(1)
Interpolation	Blocking	Non-blocking

Race Condition Prevention

Scenario	Prevention Mechanism
Two messages reach threshold simultaneously	AtomicBool::compare_exchange ensures exactly one caller wins
Message arrives during interpolation	Interpolation uses cloned snapshot; live vector safe to modify
Double interpolation trigger	Atomic flag checked before AND after threshold

Test plan

• Run batch_recon tests
• Run full test suite
• Run tests with multiple threads (RUST_TEST_THREADS=8)

🤖 Generated with Claude Code