Harden EMUAdvisor evidence and privacy boundaries

[AFR0011]

Summary

• relabels the 60-case evaluation set as assistant-curated and pending independent review while preserving protected case content and order
• replaces semantic-sounding metrics with versioned automated evidence/retrieval/behavior proxies and expected-evidence citation checks
• adds server-issued session capabilities, admin-only minimized enumeration, and default-off transcript/audit persistence
• requires explicit fixture or artifact corpus mode, visibly marks fixtures, and validates exact-host HTTPS redirects
• adds portable Windows/Ubuntu Python 3.12 CI, pinned audit tooling, and immutable official Action references
• reconciles public claims and workflow evidence without publishing corpus, logs, transcripts, or generated metrics

Verification

• independent tester: PASS_WITH_RISKS after one documentation-only repair/retest; the initial FAIL is preserved in QA_REPORT.md
• 42-file syntax scan
• 81/81 tests passed twice, including a clean Windows CPython 3.12 environment
• focused security/provenance tests: 11/11
• all three evaluation-set validations and review-status checks passed
• publication guard and bootstrap audit passed
• browser smoke passed twice after one transient timeout
• pip check passed; pip-audit found no known vulnerabilities
• protected v1_gold content/order comparison passed
• tracked-tree credential-pattern scan found no matches

Still blocked / unverified

Presentation, publication, release, deployment, and production-readiness remain blocked. Remote Ubuntu CI, live official-host crawl behavior, artifact-backed Qdrant/Ollama operation, history-aware secret review, human semantic review/institutional attestation, corpus redistribution rights, and Internet-grade identity/deployment security remain pending.

No merge, release, or deployment is requested by this PR.