Provider-agnostic sliding window rate limiter for LLM API calls. Single Go package (no sub-packages) with two persistence backends: YAML (single-process, default) and SQLite (multi-process, WAL mode). Enforces RPM, TPM, and RPD quotas per model. Ships default profiles for Gemini, OpenAI, Anthropic, and Local providers.
Module: `forge.lthn.ai/core/go-ratelimit` — Go 1.26, no CGO required.
- **No `init()` functions**, no global mutable state
- **Mutex discipline**: lock at the top of public methods, never inside helpers. Helpers that need the lock document "Caller must hold the lock". `prune()` mutates state, so even "read-only" methods that call it take the write lock. Never call a public method from another public method while holding the lock.
-`sqlite.go` — `sqliteStore` internal type, schema creation, load/save for quotas and state
Constructor matrix: `New()` / `NewWithConfig()` for YAML, `NewWithSQLite()` / `NewWithSQLiteConfig()` for SQLite. Always `defer rl.Close()` with SQLite.
### Sliding window
1-minute window pruned on every `CanSend`/`Stats`/`RecordUsage` call. Daily counter is a rolling 24h window from first request, not a calendar boundary. Empty state entries are garbage-collected by `prune()` to prevent memory leaks.
## Test Organisation
White-box tests (`package ratelimit`), all assertions via `testify` (`require` for fatal, `assert` for non-fatal). Do not use `t.Error`/`t.Fatal` directly.
SQLite tests use `_Good`/`_Bad`/`_Ugly` suffixes (happy path / expected errors / edge cases). Core tests use plain descriptive names with table-driven subtests. Use `t.TempDir()` for all file paths.
