A title request against a slow local model (a reasoning model on modest
hardware) hit `auxiliary.title_generation.timeout`, was retried on the same
provider twice more with backoff, and only then fell back — roughly four
configured windows, exactly the 3x30s the reporter read off the Ollama access
log — while the auto-title thread outlived the deadline the user thought they
had set. The failure was logged at INFO as a "connection error", the same
label as an unreachable endpoint, so the only WARNING the operator saw came
from the fallback provider complaining about a model it never had, which reads
as "your base_url was ignored".
Title generation joins compression and vision in the existing
critical-path rung that skips the same-provider retry after a full-budget
timeout, and the fallback ladder now classifies a timeout before the
connection-error rung and logs it at WARNING with the endpoint, the budget
and the config key to raise. The route carries its effective timeout so the
message can say how long it waited. Docs list the title default and the
one-window rule.
Slim redo of the retry/fallback half of #100013 on the existing rung; its
10s default and concurrency change are not carried.
Fixes#89445
Part of #66251
Co-authored-by: AJ Fasano <ajfasano@gmail.com>
f85cd964e1