Implement Stage 16 benchmarks and integration gate

[aarambh-darshan]

Summary

Implements Stage 16: benchmarks and regression test suite.

This PR adds a non-runtime benchmark crate, committed benchmark report, Stage 16 integration gate, CI benchmark smoke, and documentation updates marking Stage 16 complete.

What Changed

• Added manas-benches, a workspace tooling crate with a custom no-dependency benchmark harness.
• Added B1-B8 benchmark coverage:
  • B1 single teach
  • B2 single ask
  • B3 .manas save
  • B4 .manas load
  • B5 tokenizer on 1000 words
  • B6 anti-forgetting proof
  • B7 estimated 1000-neuron memory footprint
  • B8 brain file growth per fact
• Added BENCHMARKS.md, generated by the full benchmark command.
• Added manas-cli/tests/stage16_integration.rs with IT-1 through IT-8:
  • 22-fact neural-weight demo
  • anti-forgetting anchor survival
  • persistence save/load/save/load
  • growth behavior under novel and repeated input
  • frozen-neuron protection stress test
  • compression file shrink proof
  • stale freshness warning
  • .txt, .md, and .rs ingestion
• Updated CI to run benchmark smoke with cargo bench -p manas-benches -- --quick.
• Updated README, roadmap, architecture, and changelog for Stage 16 completion.
• Included Cargo.lock update for the new workspace crate.

Verification

Passed locally:

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

Demo Result

bash demo.sh still passes the neural-weights-only proof after deleting all historical sidecars. The demo answers the five Stage 13 questions from neural weights with confidence 1.00, and the generated brain remains under the size gate.