Module cache 

Process-wide lockfree-ish cache of loaded VelesIndexes.

Shared between the MCP and gRPC servers (and anything else that wants to serve searches from in-memory indexes without re-walking <repo>/.veles per request).

Design notes:

• Storage: DashMap<String, CacheEntry> — sharded internal locks, so concurrent operations on different repos never contend. The “lockfree” label is the practical kind: contention is bounded to a single shard, not the whole map.

• Per-index synchronization: each cache entry holds an Arc<RwLock<VelesIndex>>. Read-only operations (search, defs, refs, …) take a shared read lock; update_from_path takes the exclusive write lock. Two clients searching the same repo proceed in parallel; an update briefly blocks readers.

• Build deduplication: a OnceCell lives inside every entry, so several concurrent loaders of the same repo cooperate — one thread does the (slow) walk + embed + load, the others await its result. No wasted duplicate builds.

• LRU eviction: each entry stores an AtomicU64 last_access. Every hit / miss bumps a global counter and writes it into the entry. When the cache exceeds capacity we scan and evict the smallest. O(N) but N is small (≤ 16 in practice).

Tests assume the eviction is “eventually correct”, not strictly LRU under contention — two threads concurrently bumping last_access on different entries may finish in arbitrary order. For the actual workload (10-slot cache, ~tens of repos per session) this is fine.

Structs

IndexCache

Lockfree-ish process cache of loaded indexes.

Constants

DEFAULT_CACHE_SIZE

How many VelesIndex entries the cache keeps before evicting LRU.