baef5ba95d
PrefixFingerprint::compute is called once per turn by the turn loop prefix-stability check. The tool-side work serializes every tool to the chat-API JSON shape, sorts the resulting strings, joins with newlines, and SHA-256s the result. For a 60-tool catalog that is ~25-40 KB of allocation plus a sort, all of which produces a byte-identical output once the tool set is stable across turns (the common case after the first turn of a session). Introduce a process-local ToolCatalogCache that stores the joined+sorted catalog under a content-derived u64 identity (length + per-tool name + description + serialized input_schema). On a hit, the per-tool JSON serialization, sort, and join are skipped entirely — the pre-computed SHA-256 hex digest is returned directly. The cache lives on PrefixStabilityManager (per-session ownership) and backs a new PrefixFingerprint::compute_with_tool_cache entry point. check_and_update, PrefixStabilityManager::new, and pin() all use the cached path. The original compute() is kept as a fallback for callers that do not have a cache in hand (e.g. CLI tools that build a one-shot fingerprint). The cache is bounded (default capacity = 8) and uses insertion-order eviction, matching the eviction strategy already in transcript_cache.rs. invalidate() is exposed for tool-registry hot-reload and MCP attach paths. Tests: 8 new unit tests cover the miss/hit path (pointer-equal Arc on hit), identity collisions, schema change detection, capacity eviction, invalidate, empty slice, and the equivalence between cached and uncached fingerprints. The full 30-test prefix_cache suite passes; the wider prefix-cache contract tests in settings, prompts, and core::engine::tests continue to pass.