research: retain actual source replay and a short human review task

[LE0-Lin]

Evidence retained

• Complete the existing Next.js pinned-source probe using the unchanged frozen 0.22.0 code and protocol. A bounded curl download completed; Windows stopped before scanning on a path-case mismatch; an existing Ubuntu WSL environment with a native Linux directory passed the same checks and produced the actual report.
• Retain the original seven failed stages unchanged and record the new failed transfer, successful transfer, Windows failure and Linux replay separately.
• Record verified source membership/blobs and native (27 files) versus root-only (1 file) observations. Findings are unscored review signals, not confirmed defects. Matched AI/tool evidence and human reference labels remain absent; metrics stay null.
• Prepare a readable 28-line human preannotation task selected only by smallest retained text size. The user agreed to read it; no annotation is inferred from consent. Keep the parent corpus unresolved.

Boundaries

No scanner rules, runtime behavior, package version, published artifacts, frozen inputs, prior AI answer or human labels are changed. No new AI run, downloaded command, hook, dependency, project test or maintainer message was executed. The real probe is with frozen 0.22.0, not published 0.23.0.

Validation

276 local unittest cases completed successfully with one platform-specific skip. Repository self-scan and diff checks passed. New tests bind retained metadata and source bytes; they do not authenticate execution or establish correctness. Update the previous failure-only assertion so unchanged failed attempts can coexist with the separately recorded actual replay.