Implement Stage 18 internet refresh agent

[aarambh-darshan]

Summary

• Add the new manas-agent runtime crate for explicit internet refresh through DuckDuckGo.
• Add manas refresh, manas refresh --fast, manas refresh --dry-run, and manas refresh --limit N.
• Persist refresh metadata in .manas format v3 while keeping v2 brain files loadable.
• Re-teach refreshed facts into neural weights so manas ask remains local-only and never auto-fetches.
• Update docs, roadmap, architecture, changelog, benchmark report, and Cargo.lock for Stage 18 completion.

Implementation details

• manas-agent isolates ureq and serde_json from core learning/storage crates.
• Refresh planning scans stale hidden memories with stored memory_input / memory_target, dedupes queries, and supports fixture clients for deterministic tests.
• DuckDuckGo parsing uses AbstractText, then Answer, then RelatedTopics text.
• Trainer refresh grows fresh neural memory when the stale match is protected, preserving frozen weights.
• Bound-memory query selection can prefer a fresh refreshed duplicate over stale memory when scores are effectively tied.
• Store writes format v3 with refresh metadata and reads legacy v2 files with empty refresh fields.

Verification

• cargo fmt --all -- --check
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo build --workspace --release
• cargo bench -p manas-benches -- --quick
• cargo bench -p manas-benches -- --write-markdown BENCHMARKS.md
• bash demo.sh
• git diff --check