Expose the Bedrock inference-profile id as a workflow_dispatch input
(falling back to the E2E_AWS_BEDROCK_MODEL repo variable, then the test's
Haiku 4.5 default) so a manual run can drive the provider matrix with
Sonnet 4.6 — the model from the original cost report — without a code
change. The account-enabled profile id is deployment-specific, so it is
supplied at dispatch time rather than hardcoded.