[nicoloboschi]

Closes #5561

Problem

When the worker's wall-clock limit cancels a knowledge-page refresh, the engine records the failure and always pauses automatic refresh. If the limit fired while the LLM call was waiting out a retry backoff (provider down: DNS failure, 5xx), the page stayed paused for days after the provider came back. #5399 already skips the pause for those same temporary errors when they surface inside the limit.

Fix

• The worker passes the stage it was in at the timeout (already in the log line) to the engine's timeout hook.
• on_task_wall_timeout still records the failure in the page's history, but skips the pause when the stage ends in .backoff.
• A timeout during a running LLM call still pauses (keeps Unbounded paid-LLM spend: mental-model refresh loop never converges (no cost/rate guard) #4532's protection).

Tests

• New test_refresh_cut_off_during_a_retry_backoff_does_not_pause: not paused, one failure in history, the scheduler queues it again. Fails on the old code.
• test_refresh_cut_off_by_the_wall_timeout_counts_as_failed now passes a running-call stage and still expects a pause.
• Worker notification test checks the stage is passed through.

No system-test story: the path needs a real wall-clock timeout during an LLM backoff, which the blackbox stubs can't drive in reasonable time; the mechanism is covered at the engine + worker level.