Preserve planning artifacts and harden training telemetry

[vikranthreddimasu]

Summary of improvements

• Upload a product-improvement-report artifact even for duplicate daylight-saving fallback schedule runs, so automation can inspect the newest run without artifact lookup failures.
• Add fail-fast validation for TrainingConfig values that would otherwise produce silent zero-step or invalid training runs.
• Stop recording throughput metrics for failed step contexts, keeping scheduler/monitoring signals clean after exceptions.

Security changes

• No auth, authorization, credential handling, secrets, or destructive-action code changed.
• Diff-only secret scan returned no matches.

Usability changes

• Product-improvement automation now gets a clear duplicate-schedule report instead of a successful run with no artifact.
• Invalid training-loop options now raise direct TrainingConfigError messages.

Tests/checks run

• python -m pytest tests/test_utils/test_source_hygiene.py -q
• python -m pytest tests/test_training/test_loop.py -q
• python -m pytest tests/test_integration/test_end_to_end.py::TestTrainingLoop::test_single_node_training_loop -q
• python -m pytest tests/test_monitoring/test_throughput.py -q
• make lint
• python -m pytest tests/ -q (1002 passed, 2 warnings)
• GitHub Actions test matrix: all 11 checks passed
• git diff --check master...HEAD before merge
• Diff-only secret scan with rg

Commit list

• 36fbf7d Preserve planning reports for duplicate schedules
• 94b2377 Validate training loop config
• 1ca1a9c Skip throughput metrics for failed steps

Risks or manual review notes

• Changes include one GitHub Actions workflow tweak. It keeps contents: read permissions and only changes reporting behavior for skipped duplicate schedule runs.
• No high-risk auth/security/credential path changed.

Whether the PR was merged

• Merged into master on 2026-07-16 as squash merge commit 4c36722.