Summary

• preserve an explicit max_wait_duration=0 when building custom-training scheduling inputs
• continue treating None as an omitted duration
• add regression coverage for omission, zero with Flex Start, zero without another scheduling option, and a positive duration

Why

The documented value 0 is a sentinel that requests an indefinite Dynamic Workload Scheduler wait. Request preparation used truthiness checks, causing explicit zero to be converted to None and silently omitted from the outgoing training pipeline.

Using explicit is not None checks keeps the API distinction between omission and zero without changing positive-duration behavior.

Fixes #7067.

Validation

• focused regression test: 4 passed, 236 deselected
• complete tests/unit/aiplatform/test_training_jobs.py module: 240 passed
• negative control: restoring the original truthiness checks makes both zero cases fail with None != "0s"
• Black 24.8.0 check passes for the changed production file
• Flake8 6.1.0 passes for both changed files
• git diff --check passes

A broader local Windows package run was not completed because the Python process encountered a native memory-access failure; no full-suite result is claimed. GitHub's Linux CI remains the broader verification environment.