docs: add categorized PR analysis replays

[zhao-stack]

What changed

• Add raw, categorized vLLM PR replay outputs for three actionable-break cases and one verified no-break case.
• Add two known scope-limitation cases so a narrow PR-gate PASS is not misread as proof of runtime compatibility.
• Add a machine-readable manifest, SHA-256 checksums, and a human/AI reading guide.
• Link the bundle from the main README.
• Add regression tests that keep the manifest, report metadata, classifications, and checksums synchronized.

Why

Users and AI agents need concrete examples of both BREAKS FOUND and PASS, including the exact SHA and analyzer-version context. The separate scope-limitation category documents that the vllm-interface scenario intentionally skips monkey patches and cannot be interpreted as an all-runtime guarantee.

User impact

Readers can now compare JSON, CSV, Markdown, and metadata outputs for real PR ranges and use manifest.json as a stable machine entry point.

Checks

• python -m ruff check tools tests
• python -m pytest tests/e2e/vllm_interface/generator_tests/test_pr_analysis_results.py (2 passed)
• Manifest-to-report metadata validation for all 6 cases
• SHA-256 verification for all 24 archived report artifacts
• Staged-index checksum replay and public-content credential/path scan