Fix Stage 13 answer decoding noise

[aarambh-darshan]

Summary

• pack meaningful answer word IDs directly into neural output weights for Stage 13 bound-memory facts
• decode packed answer IDs before falling back to legacy dense-vector matching
• make the legacy fallback conservative so old brains avoid unrelated filler words like 'lisa'
• tighten demo.sh and the Stage 13 integration test to reject unexpected answer-line words
• update docs to describe packed answer-word decoding instead of loose nearest-token decoding

Behavior

• existing old-format root brain now answers 'animal' for 'what is a cat?' instead of 'animal lisa'
• fresh Stage 13 demo brains answer 'small domesticated animal fur whiskers'
• all demo answers still come from neural weights after sidecars are deleted
• brain.manas remains 82716 bytes in the demo, below the 500KB gate

Verification

• ./target/release/manas ask "what is a cat?"
• ./demo.sh
• cargo fmt --all -- --check
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• cargo build --workspace --release