4.12.0

4.12.0

• Keep one in-flight poll per task family while filling the configured workflow
  and activity execution slots (#199). A received task finishes hydration and
  initial lease confirmation before polling refills a spare slot. Its execution
  slot stays occupied through completion reporting, so slow callbacks can run
  concurrently without multiplying idle long polls.

• Reuse worker-authenticated activity lease capability observations instead of
  fetching /cluster/info for every activity (#199). Client clones share one
  discovery when needed and refresh only after an explicit capability refusal. Original claim,
  ownership and deadline validation still apply to every renewal.

• Add a public raw request API for tools that must reach every Server route
  and report its response unchanged (#180). Client::control_plane_request
  and Client::worker_plane_request take a RawRequest (path segments the
  SDK percent-encodes, query pairs, a verbatim JSON body, a timeout and
  optional version-header overrides) and return a RawResponse for every
  HTTP status, with the exact body bytes and all headers.
  RawResponse::error_for_status gives the same Error::Http or
  Error::Protocol a typed call would, and refusal() decodes reason,
  rejection_reason, message, error and validation_errors. Raw
  requests make one attempt and never rewrite bodies. Worker-plane 2xx
  responses must echo a same-major, equal-or-newer-minor protocol version in
  both the header and the body. Typed methods remain the API for
  applications.

• Add ClientBuilder::danger_accept_invalid_certs for self-hosted Servers
  whose certificates cannot be verified. It defaults to false; prefer
  additional_root_certificates_pem for a private CA.

• Replay condition-wait histories in which the Server recorded a second open
  of a wait that was still open (#191, #192). A wait is now linked to the
  same logical wait by its state when the next open was recorded: if it was
  still open, the next open is a repeated open. So an orphaned duplicate whose
  timer fires later no longer fails replay with
  condition_wait_reopened_after_timeout (the sdk-php#283 shape). On PHP
  Server 2.x, condition members of one parallel or select group woken by
  one signal no longer fail the next task with
  condition_wait_group_reopen_mismatch, and each member consumes its own
  reopen, including when another member such as a timer sits between them
  (#194). A pending scope cancellation over such a group selects its original
  members by sequence. A wait that timed out before the next open still ends
  it. Both histories were refused by 4.11.1 as well.