CON-1516: Remove vCPU gate from self-test preflight

[jjziets]

Summary

• remove the self-test preflight hard gate based on offer["cpu_cores"] >= 2 * num_gpus
• leave physical CPU validation to the self-test runtime image, which can inspect visible physical cores
• reduce noisy CLI-side self-test requirement output by removing the old vCPU/core preflight diagnostic from the local preflight report
• map 5001/udp when launching the self-test image and externally probe it after TCP /progress is reachable
• add distinct UDP diagnostics for missing 5001/udp mapping and for “TCP works, UDP echo failed”
• update the legacy vast.py copy and regression coverage

Scope Note

• This PR is the CLI side of physical CPU-core preflight alignment plus UDP self-test verification.
• The paired self-test image responder is in vast-ai/self-test#6. The CLI expects that image to echo UDP probes on 5001/udp.

Validation

• ./.venv/bin/python -m py_compile vast.py vastai/cli/commands/machines.py vastai/cli/self_test/runtime_diagnostics.py
• ./.venv/bin/python -m pytest tests/cli/test_machines_commands.py -q — 51 passed, 1 existing pytest config warning

Dogfood

• Paired self-test image staging build was triggered from vast-ai/self-test#6 with tag prefix self-test-udp-dogfood-cuda-.
• Image run: https://github.com/vast-ai/self-test/actions/runs/28094691046
• Validation jobs passed for CUDA 11.8, 12.8, 13.0, and 13.3. Image build/push jobs are still in progress as of the PR update.
• Once the staging tag is available, the host-verification dogfood command should use --test-image vastai/test:self-test-udp-dogfood-cuda-12.8 or the matching CUDA tag.
• The local client-image dogfood harness has also been updated to request/probe 5001/udp for paid client rentals, but that is not host verification.

[robballantyne]

Looks good. I will approve without requesting changes when test image is available

[jjziets]

CON-1531 follow-up pushed in 63f2752.

What changed:

• Broadened self-test progress-stage parsing for the current image output (ResNet18 test on all GPUs, ECC test on all GPUs, NCCL distributed test with N GPUs, stress-ng and gpu-burn tests simultaneously).
• Improved the offline runtime summary so it covers instances going offline before or during runtime.
• Split repeated status polling/API failures into instance_status_poll_failed instead of folding them into startup timeout.
• Cleaned up startup-timeout wording for hosts.
• Surfaced cleanup_failed if the workload passes but the temporary paid instance cannot be destroyed.
• Added tests for current image stage lines, status-poll timeout classification, and cleanup failure.

Validation:

• ./.venv/bin/python -m pytest tests/cli -q -> 313 passed, 1 existing config warning.
• Paid dogfood with latest PR image vastai/test:self-test-udp-dogfood-cuda-12.8:
  • Green machine 35008: passed TCP, UDP, system requirements, ResNet18, ECC, NCCL, stress/gpu-burn; cleanup succeeded.
  • Red machine 141071: preflight caught low reliability/upload; forced runtime failed with progress_endpoint_unreachable; support bundle printed and cleanup succeeded.

Local evidence/artifacts:

• /Users/hanneszietsman/VastAi/CON-1531-self-test-error-flow-map.md
• /Users/hanneszietsman/VastAi/dogfood-captures/host-self-test-latest-pr/20260625T090214Z.tar.gz

[jjziets]

Final CON-1531 diagnostic follow-up SHA is now f737010 after amending in the catch-all unexpected_error path.

Additional tightening in the final amend:

• Added first-class unexpected_error catalog copy and runtime diagnostic rendering for otherwise unhandled CLI exceptions.
• Added preflight_checks stage marking before threshold checks run.
• Added regression coverage for unexpected exception shaping/redaction.

Latest local validation:

• ./.venv/bin/python -m pytest tests/cli/test_runtime_diagnostics.py tests/cli/test_self_test_support_bundle.py tests/cli/test_machines_commands.py -q -> 84 passed, 1 existing config warning.
• ./.venv/bin/python -m pytest tests/cli -q -> 314 passed, 1 existing config warning.