fix(agent): keep standalone greetings tool-free

[ZhiXiao-Lin]

Summary

• recognize only exact standalone conversational greetings after whitespace and terminal-punctuation normalization
• expose no tools for those turns, preventing an initial greeting from starting workspace scans
• suppress synthetic continuation injection for the same prompt so a friendly response such as I'll be happy to help... is terminal
• preserve the normal tool surface when a greeting also contains an action request
• document the behavior and add English/Chinese regression coverage

Scope

This intentionally fixes the concrete standalone-greeting path without trying to guess that every ambiguous prompt is non-actionable. Hello! Inspect this repository... still receives the normal tools.

Related to #32. This PR does not close that issue; broader ambiguous-input behavior still needs reporter confirmation with a current version and trace.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --lib --bins -- -D warnings
• cargo test --workspace
• cargo test --workspace --all-features --lib
• cargo run -p a3s-code-core --example agent_convergence_benchmark (4/4 cases passed)
• node scripts/generate_event_protocol_artifacts.mjs --check
• node scripts/sdk_api_alignment_check.mjs
• cargo check --manifest-path sdk/node/Cargo.toml
• cargo check --manifest-path sdk/python/Cargo.toml