feat: complete context, security, and release evidence

[DuangZ-GR]

Summary

• add structured context state, provider usage/smoke gates, shared parent/subagent budgets and absolute deadlines
• add native Ollama bridge mode and traceable agent/context benchmark artifacts
• harden symlink traversal, web-search argument handling, network permission gates, and tool resource limits
• add security heldout benchmark, doctor JSON diagnostics, install/demo scripts, Linux/Windows CI and release dry-run workflow
• aggregate 13 benchmark evidence sources with source SHA-256 and claim-eligibility boundaries

Why

M17/M18 close the project's context/agent evaluation and release-engineering gaps while preventing smoke or unsupported cache data from being presented as formal results.

Validation

• cargo fmt --all -- --check
• cargo clippy --locked --all-targets -- -D warnings
• cargo test --locked --all - 177/177
• PYTHONPATH=python python -m pytest -q python - 350/350 in zgr
• frozen Linux security heldout: 12/12 applicable attacks intercepted/gated; 0/8 benign false blocks; 3 explicitly not applicable
• offline demo preserved source SHA-256
• final JSON/Markdown evidence regenerated byte-for-byte

Boundaries

• qwen2:0.5b smoke is claim-ineligible; no formal multi-agent/provider benefit claim
• provider cache hit rate remains null/unsupported
• GitHub Release/tag is not created by this PR
• Windows junction behavior awaits hosted Windows CI