Fix missing task.duration metric for tasks that get skipped after starting.

[myps6415]

Summary

task.duration (and its registry-derived legacy name dag.<dag_id>.<task_id>.duration) was emitted for tasks ending in SUCCESS or FAILED but missing entirely for SKIPPED tasks, despite the metric being documented as available for all terminal states. Reproducible with ShortCircuitOperator, BranchPythonOperator, and BashOperator with skip_on_exit_code (narrowed in #61849 by @shivaam).

Fix

The metric is emitted from finalize() in the Task SDK task_runner, gated on if ti.start_date and ti.end_date:. The SUCCESS and FAILED exception handlers set ti.end_date on the local RuntimeTaskInstance before constructing the outbound TaskState message, so the guard passes. The two SKIPPED handlers (AirflowSkipException, and the SKIPPED branch of DagRunTriggerException with skip_when_already_exists=True) set end_date only on the outbound TaskState message, leaving ti.end_date as None — so finalize()'s guard failed for skipped tasks.

Set ti.end_date on the local instance in both handlers, mirroring the AirflowFailException and _handle_current_task_success patterns. The TaskState message references ti.end_date to keep the local instance and the outbound message in sync.

Test plan

• New parametrized test_task_duration_metric_emitted_for_terminal_states covers success / skipped / failed — verifies stats.timing("task.duration", ...) is called with the correct tags and that the legacy dotted form is also emitted via the metrics_template.yaml registry.
• Self-verified by stashing the production change: the [skipped] parametrize case fails without the fix and passes with it; [success] and [failed] pass in both cases.
• Existing SKIPPED-related tests (test_run_basic_skipped, test_task_runner_calls_listeners_skipped, etc.) still pass — no regression.
• ruff format / ruff check / mypy-task-sdk / prek run --from-ref upstream/main --stage pre-commit all green.

Note: test_handle_trigger_dag_run_conflict[True-skipped] was already failing on upstream/main due to a mock-assertion mismatch (the test uses mock.call.send(msg=TriggerDagRun(...)) as kwarg but _handle_trigger_dag_run calls SUPERVISOR_COMMS.send(TriggerDagRun(...)) positionally). Unrelated to this PR.

closes: #61849

---

Was generative AI tooling used to co-author this PR?

• Yes — Claude Code (Opus 4.7)

Generated-by: Claude Code (Opus 4.7) following the guidelines