Add resilient AI service handling

docs/ARCHITECTURE.md

@@ -49,6 +49,9 @@ questions without applying a project update.
 All structured AI updates pass through a scope-aware merger. It applies only
 the scopes declared by the response, ignores user-locked scopes and preserves
 manual control values.
+AI provider calls can be wrapped by a configurable retry service. It retries only
+temporary network, server and rate-limit failures, honors a provider-supplied
+retry delay when available, and propagates cancellation without retrying.
 
 ## Prompt Compiler
 

docs/TASKS.md

@@ -63,7 +63,7 @@ requirement is missing and blocks implementation, record it in
     and production decisions.
 -   [x] Implement optional Discuss mode.
 -   [x] Enforce user-lock/manual-value precedence over AI output.
--   [ ] Add error, retry, cancellation and rate-limit handling.
+-   [x] Add error, retry, cancellation and rate-limit handling.
 
 ## Phase 5 --- Arabic Lyrics Processing