fix(e2e): avoid parsing stage logs as final artifact paths

[fangyangci]

Summary

• Extract deployment artifacts only from Final artifact: or Existing artifact found:, preferring the final artifact over reused or intermediate outputs.
• Preserve ANSI stripping and Rich-wrapped paths, separate captured stderr/stdout, and handle OSError/ValueError when checking candidate files.
• Add regression coverage for stage output, invalid paths, cache fallback, output-stream boundaries, and final-artifact precedence. Apply the configured formatting hooks to the two touched files.

Root Cause

The wrapped-path reconstruction introduced in #1411 also parsed stage-level Artifact: lines. Those lines include a size suffix, such as ..._export.onnx (335.3 MB), so the .endswith(".onnx") stop condition never matched. Subsequent warning and Optimize output was appended to the candidate filename. On the affected Windows agent, Path.is_file() raised WinError 1113 after the build subprocess had already exited successfully.

Stage artifacts are not necessarily the deployment-ready model and must not be selected as a fallback for the final build summary.

Validation

• Complete E2E runner unit tests after formatting: 364 passed, including 15 artifact-path cases.
• Configured pre-commit hooks passed for both changed files; editor diagnostics and git diff --check passed.
• Real local Windows E2E reruns of openai/clip-vit-base-patch32 with QNN package 2.2480.49.0, matching the version reported in the failing logs:

Task	Target	Runner job labels	Result
feature-extraction	QNN NPU	fp32, w8a16	2/2 PASS; runner exit 0
zero-shot-image-classification	QNN NPU	w8a8, w8a16	2/2 PASS; runner exit 0
zero-shot-image-classification	QNN GPU	default	1/1 PASS; runner exit 0

All three runs used --eval-type perf --timeout 1800 --raw-output with isolated temporary output/cache directories. Builds exercised the stage-output parsing path, and all eight component perf subprocesses exited 0. No path-parser exceptions occurred. Saved result JSON confirms all five jobs passed.

This validates the perf workflows from the three reported failures; accuracy was not rerun. The portable unit regression uses actual Rich output and injects the affected Windows filesystem exception; the hardware E2E reruns use real build/perf subprocesses. Temporary models, raw logs, and historical evaluation results are not part of this PR.