AI Implementation feature(1089): Scheduled Execution with Cron and Retry Handling (#9)

docs/architecture/pipeline-runner.md

@@ -97,15 +97,12 @@ characters); the runner fails safely (raises) rather than truncating.
 
 # Retry Behaviour (`_run_with_retry`)
 
-* `attempts = max(1, config.max_retries + 1)` — at least one attempt
-  even if `MAX_RETRIES=0`.
-* Any exception inside `_run_once` is caught (`noqa: BLE001` —
-  intentional, the whole pass is opaque to the runner).
-* Each failed attempt is recorded via
-  `log_error("pipeline_error", exc=exc, attempt=attempt, max_attempts=attempts)`.
-* When the budget is exhausted, the runner emits
-  `log_error("pipeline_failed", exc=last_exc, attempts=attempts)` and
-  returns `None` so `main` can translate it to `exit 1`.
+* `attempts = max(1, config.max_retries)` — at least one attempt even if `MAX_RETRIES=0`.
+* The runner owns the unified retry budget; `_run_once` calls `ensure_repo` with `max_retries=0`.
+* Transient Git and Mastodon failures are retried with delays of `2 ** (attempt - 1)` seconds between attempts.
+* Each failed attempt emits `retry_attempt` with `operation`, `attempt`, and the exception `error` text.
+* Fatal configuration, local repository, and publishing validation errors emit `retry_exhausted` immediately without retrying.
+* When the budget is exhausted, `retry_exhausted` includes the final operation, total attempts, and error text, and `main` returns exit code `1`.
 
 # Silent-on-No-Matches Contract