Add staged CI auto-fix workflow and production safeguards

[Kate-M1125]

What changed

• add the complete staged CI repair flow: diagnosis, patch generation, static validation, isolated tests, independent review, and Draft PR publishing
• lock patches across authorization stages with SHA-256 and reject tampered or unreviewed artifacts
• add explicit RUN_PATCH and CREATE_DRAFT_PR confirmations, task ownership, Bearer API keys, concurrency limits, and append-only stage auditing
• persist API jobs in SQLite and recover interrupted service state
• make GitHub publishing idempotent by recovering verified branches and existing Draft PRs
• add macOS and Linux network-denied test sandbox support with resource limits and secret-stripped environments
• add Web UI support for Actions Run URLs, patch/review results, authorization actions, and audit history
• split core and optional RAG dependencies, remove import-time embedding initialization, and add Docker deployment files
• add a GitHub Actions unit-test workflow

Why

RepoMind previously stopped at read-only CI diagnosis. This completes the workflow while preserving explicit human authorization and hard safety gates before executing repository code or writing to GitHub.

Safety boundaries

• no patch execution without explicit confirmation
• no GitHub write without a second explicit confirmation
• only the exact SHA-256-locked, successfully tested patch can be published
• high/critical review findings, fork ambiguity, missing review verdicts, and invalid patches block publishing
• generated PRs are always Draft and are never merged automatically
• duplicate publication requests recover existing resources instead of overwriting branches

Validation

• 255 local non-regression tests passed; 48 environment-dependent tests skipped
• Python compile, shell syntax, workflow YAML, Docker Compose config, and git diff checks passed
• GitHub Actions push run 31896378642 passed
• GitHub Actions pull_request run 31896381388 passed
• code review found and fixed an unescaped label rendering path and a permissive missing-review publish gate

Full model-quality regression tests remain opt-in because they require real API credentials, cloned fixture repositories, and external model calls.