Detect Docker host-control exposure in agent configs

[brandonwise]

Summary

• Add AW-013 to detect Docker socket exposure and privileged Docker host controls in MCP/agent configs.
• Add a Docker host-control fixture plus unit and integration coverage for critical socket exposure and high-risk privileged flags.
• Refresh vulnerable Rust transitive dependencies in Cargo.lock so cargo audit is clean again.
• Harden baseline unit-test temp directories after the merge exposed a parallel-test collision.
• Update the README detection rule table and trust-signal test count.

Rationale

Current telemetry is clustered around agent trust boundaries, sandbox escapes, and production agent consequences:

• TechRadar covered recent AI coding-agent sandbox escape issues: https://www.techradar.com/pro/security/top-ai-coding-agents-can-be-easy-victims-to-sandbox-escapes-showing-they-arent-as-secure-as-they-claim-to-be
• Product Hunt AI Agents discussion frames production agents around action consequences, not just answer quality: https://www.producthunt.com/categories/ai-agents
• The public AI-agent incident/timeline space is active: https://github.com/webpro255/awesome-ai-agent-attacks

Validation

Local checks:

• python3 skills/tiktok-trends/scripts/research.py -> PASS; saved projects/last-tiktok-trends.md.
• cargo fmt --check -> PASS.
• cargo test docker -- --nocapture -> PASS; 5 Docker unit tests and 1 integration test passed.
• cargo test -> PASS; 230 unit tests and 43 integration tests passed.
• cargo clippy --all-targets -- -D warnings -> PASS.
• cargo build --release -> PASS.
• target/release/agentwise scan testdata/docker-host-control.json --format json -> PASS; emitted 1 critical AW-013 Docker socket finding and 1 high AW-013 privileged Docker finding.
• cargo audit -> PASS; no vulnerabilities reported across 192 crate dependencies after lockfile refresh.
• From projects/humanizer: node src/cli.js score -f ../agentwise/README.md --ignore-code -> PASS; README scored 41/100.

GitHub checks:

• Format -> PASS.
• Clippy -> PASS.
• Test -> PASS.
• Action smoke -> PASS.
• Build matrix -> PASS for x86_64/aarch64 Linux, x86_64/aarch64 macOS, and x86_64 Windows.

Risk / Rollback

• Low-risk additive scanner rule plus patch-level lockfile security updates.
• Rollback by reverting this PR if AW-013 creates unexpected false positives.