Use one longer deadline to classify a timeout

A timeout increase is a diagnostic discriminator, not a performance fix. Raise one aligned deadline once under rollback; if the operation consumes the new budget, stop repeating the trial and profile or remove concrete work. Do not leave the higher limit enabled from this evidence alone.

Then reproduce the production query shape with synthetic data, measure branch-level execution, and gate optional widening before it starts. Preserve positive controls that require each wider branch plus result, authorization, limit, and audit equivalence.

A local branch win remains a qualified hypothesis until one ordinary-path smoke succeeds under unchanged deadlines. The one-time trial does not identify the slow subplan or prove capacity.