Small Harness v0.4.0 speeds local LLM workflows

Local LLM agents are transitioning from neat CLI toys to robust, structured developer tools by adopting parallel execution and safety constraints directly in the harness.

• –Concurrency Boost: Running read-only actions in parallel using `join_all` dramatically decreases step latency, which is a massive win for locally hosted models where raw inference speed can be a bottleneck.
• –Safety & Verification First: Self-verification of edits (`file_edit` re-reading and confirming matches) reduces the likelihood of hallucinated code updates failing silently.
• –Context Management via Subagents: Isolating detailed investigations inside a lightweight, read-only 12-step subagent prevents local context windows from being overwhelmed by long file read traces.
• –Git-less Branching: The new native session paths allow developer experimentation with codebases without dealing with git worktree clutter.