Implement Stage 19 language generation

[aarambh-darshan]

Summary

• Add the new manas-language runtime crate for local fluent generation on top of associative-memory query results.
• Add manas generate <prompt> and manas generate --max-words N for explicit sentence generation.
• Add manas ask --fluent <question> while keeping default manas ask compact and backward-compatible.
• Generate fluent sentences from neural-weight answer concepts without using sidecars or internet fallback.
• Update docs, roadmap, architecture, changelog, benchmark report, and Cargo.lock for Stage 19 completion.

Implementation details

• manas-language keeps fluent generation separate from the core associative-memory engine.
• Generation runs after manas-learn answers from neural weights, then realizes decoded concepts into a sentence.
• Added configurable compact and expanded decoding so default ask stays compact while generation can use richer concepts.
• manas generate uses expanded query behavior for fluent output from learned neural concepts.
• manas ask --fluent provides optional sentence output without changing the default retrieval path.
• Added deterministic prompt-intent realization for definition, location, time, action, and fallback prompts.
• --max-words N limits generated sentence length for CLI-controlled output.
• No .manas format change is required; generation is deterministic, local, and built on existing persisted neural memory.
• Stage 19 tests prove generation still works after historical sidecars are deleted and still reports neural weights as the answer source.
• Added B9 benchmark coverage for single fluent generation.

Verification

• cargo fmt --all -- --check
• cargo test --workspace
• cargo test -p manas-language
• cargo test -p manas-cli stage19
• 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