Stage 15: compression and forget command

[aarambh-darshan]

Summary

• Add safe Stage 15 compression planning and execution through manas-learn::compression.
• Add Network::merge_remove_hidden_neurons so hidden neurons and output columns compact atomically.
• Add manas forget, manas forget --dry-run, and manas forget --threshold N.
• Update README, roadmap, architecture, and changelog for Stage 15 completion.

Safety model

• Only stale low-importance Open hidden neurons are candidates.
• Candidates must have a highly similar retained merge target.
• Frozen neurons and frozen output edges are never modified.
• Dry-run uses the same planning API as real compression and does not save.

Tests

• cargo fmt --all -- --check
• cargo test -p manas-core compression
• cargo test -p manas-learn compression
• cargo test -p manas-cli forget
• cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• ./demo.sh

Demo result

./demo.sh passes. The 22-fact neural-weight demo answers all five checks from neural weights and keeps brain.manas at 82716 bytes.