Historical pre-upgrade audit record: runtime, security, CI, deployment, and evidence gaps

[CoreyLeath-code]

Historical record: This issue documents problems recorded before the repository's audit/upgrade work. Checked items mean a merged PR reports the item remediated; they are not a substitute for re-validating the current default branch. The scope is the evidence in linked commits and PRs—not a claim that every historical defect has been discovered.

Baseline and remediation timeline

Earliest documented pre-upgrade baseline: d5f059c.

Repository and deployment

• No coherent local run path; README referenced nonexistent directories.
• CI workflow filename was malformed.
• Compose was incomplete and lacked a valid top-level service stack.
• Most services had no Dockerfiles, health checks, dependency ordering, or local database schema.
• Render configuration was a stub.
• Docker build context lacked a .dockerignore, risking inclusion of VCS data, environment files, Terraform state, and unrelated artifacts.

Remediated in PR #1 and PR #4.

Correctness, security, and runtime

• API modules imported a nonexistent core.* namespace.
• Tests imported the wrong application and failed during collection.
• Authentication used literal admin/admin credentials.
• Backend forced CUDA and could crash on CPU-only hosts.
• MLflow work and heavyweight model loading occurred at import/startup.
• An LLM route loaded synchronously and lacked a safe CPU/error path.
• Database-write exceptions were silently swallowed.
• Required API/test dependencies were absent, and importing the API required PyTorch.

Remediated in PR #2 and PR #3.

CI and research evidence

• Pytest invocation depended on a user-local binary being on PATH.
• README badges referenced the former repository owner.
• README mixed targets and one-off smoke measurements with insufficient claim boundaries.
• No deterministic, versioned PSI/KS benchmark artifact or enforced quality/schema regression gate existed.
• Benchmark reporting omitted a consolidated latency distribution, throughput, memory, classification metrics, provenance, and limitations.

Remediated in PR #4, PR #5, and PR #6.

Follow-up verification

• Re-run the full polyglot stack and CI on current main.
• Confirm secrets/templates remain excluded from images.
• Confirm current badges and benchmark links use CoreyLeath-code/SentinelAI.
• Reassess the still-limited focused API coverage recorded during the documentation audit.