fix: schedule lifecycle hooks for non-leaf DAG tasks on failure. Fixes #16057

[fcolombo7]

Fixes #16057

Motivation

When a non-leaf DAG task (one whose template is itself a DAG or steps) fails, its lifecycle hooks are never scheduled if the failure causes downstream tasks to be marked Omitted.

The root cause: hook evaluation in executeDAGTask runs at the top of the function while the node is still Running. For non-leaf tasks, the phase transition Running → Failed happens later inside executeTemplate. By then hook evaluation has already passed and no further sweep is triggered for that task, so the hook is permanently lost.
Leaf tasks (bare container pods) are unaffected because their phase is updated by podReconciliation before executeDAGTask runs.

Modifications

Merged the post-executeTemplate hook re-evaluation into the existing if node.Completed() block at the bottom of executeDAGTask's inner loop. Hooks are evaluated first; if not yet complete the function returns early, blocking the Omit cascade via the existing CheckAllHooksFullfilled gate in evaluateDependsLogic.
Once fulfilled, runOnExitNode proceeds as before.

Verification

Added two regression tests to dag_test.go using nested inner DAG templates (required to reproduce; bare container tasks are reconciled before executeDAGTask runs):

• TestDAGHookNonLeafSerialHeadFails: non-leaf task fails with one downstream dependent
• TestDAGHookNonLeafFanInBothFail: two parallel non-leaf tasks both fail with a shared downstream

go test ./workflow/controller/ -run TestDAGHookNonLeaf passes. Full go test ./workflow/controller/ passes with no regressions.

Documentation

No documentation change needed.

AI

Claude (Anthropic) was used to navigate and understand the codebase, identify the root cause, and assist in writing the fix and regression tests. All code was reviewed and validated by the author per the Argo project Generative AI policy.