docs(data): define configuration vulnerability annotations

[Du-professor]

Closes #8

Summary

This PR defines a consistent annotation model for configuration and non-taint
vulnerabilities and applies it to the four OpenClaw Dockerfile samples
identified by Issue #8.

The work is intentionally independent from the separate code-location-repair work. It was created
from Tencent/VulnGym:main in a separate worktree and branch, contains no
commits from that work, and changes only:

• config_vuln_annotation.md
• config_vuln_diff.csv
• data/entries.jsonl

Annotation model

• For an explicit dangerous configuration, entry_point anchors the
  configuration block, critical_operation identifies the narrow dangerous
  directive, and trace contains only meaningful state transitions.
• For a missing safety configuration, entry_point anchors the inspected
  configuration scope and critical_operation anchors runtime activation or
  the explicit end of the scope.
• trace may be empty and is not populated with routine instructions merely
  to imitate a taint flow.
• Descriptions distinguish contextual evidence, root cause, and activation,
  and do not overstate container-root impact as guaranteed host compromise.

The document also proposes optional, backward-compatible annotation_type and
missing_condition fields for future discussion. This PR does not change
SCHEMA.md or add those fields to the dataset.

Sample corrections

All source evidence was checked against OpenClaw commit
c56fb7f353d63d6ea97028ee7d8a97bc4edf21c1, referenced by
GHSA-W7J5-J98M-W679.

• entry-00241: retains FROM as the user-configuration scope anchor and
  CMD ["bash"] as the root runtime activation point; removes seven ordinary
  build instructions from trace.
• entry-00242: uses the final explicit RUN as end-of-scope evidence and
  states that the actual defect is the absent USER directive; removes the
  pseudo-taint trace.
• entry-00243: connects the inherited root user directly to ENTRYPOINT and
  removes package/copy/chmod steps that do not alter the user configuration.
• entry-00244: corrects the factual model. The container starts as app, not
  directly as root; the core defect is NOPASSWD:ALL, with a short trace
  showing USER app, demonstrated sudo capability, and runtime exposure.

verify remains 0, and all non-node fields are unchanged.

Field-level diff

config_vuln_diff.csv contains 39 auditable rows with this interface:

entry_id,node_path,change_type,field,old_value,new_value,reason

Scalar replacements are recorded per field. Trace additions and deletions
include the complete node JSON.

Verification

• JSONL rows: 408
• Changed entries: exactly entry-00241 through entry-00244
• Fixed-commit {file, line, code} checks: passed for every resulting node
• Schema and ordering invariants: passed
• Non-target fields and other 404 entries unchanged: passed
• CSV-to-JSON change reconciliation: 39/39 rows passed
• git diff --check: passed
• PR file whitelist: exactly the three files listed above
• Commit 418553f from the separate code-location-repair branch is not an ancestor of this branch

The OpenClaw source snapshot, caches, temporary validation scripts, and
credentials are not committed.