v0.2.4

Added

• Pre/post tool-call hooks on both transports. Pass pre_tool_hooks /
  post_tool_hooks to create_app(...) and build_server(...); the same
  hook fires for the same tool name across HTTP and native stdio. Pre-hooks
  raising OperationError abort the call with the transport-native error
  shape; post-hooks are best-effort (exceptions logged, never raised). New
  module app/tool_hooks.py documents the contract — HookSet,
  PreHookFn, PostHookFn, wrap_stdio_tool.
• health_extra extension point on /health/ready. Pass an async
  callable to create_app(health_extra=...); its dict is merged into the
  response body. Any returned key suffixed _ready (or the bare key
  ready) whose value is False flips the response to 503 with
  status="degraded". Probes never crash: a buggy health_extra is caught
  and surfaced as health_extra_error: true.
• metrics_replace flag on create_app(...). When true, the
  in-process metrics middleware is skipped so a downstream consumer can
  install Prometheus / OTel instrumentation without double-counting
  requests.

Changed

• app.rate_limits now declares an explicit __all__ so the per-tier
  Depends helpers (limit_local, limit_standard, limit_document,
  limit_heavy) are part of the documented public API. Renaming or
  removing a tier from now on requires a MINOR bump.
• app/stdio_server.py _serialize_errors replaced by
  app.tool_hooks.wrap_stdio_tool. Behaviour for callers without hooks is
  identical: OperationError → dict; everything else propagates.

Tests

• 63/63 passing (was 48). New tests/test_tool_hooks.py covers both
  success and error paths for every new surface: pre-hook abort,
  post-hook success/exception swallowing, health_extra 200 / 503 /
  buggy-callable, metrics_replace on/off, public re-exports.