fix(coding-agent): recover command failures and preserve session state

[MelodyVAR]

Foreground commands can stall a non-interactive session when a child process keeps an output pipe open, and shortened error output can hide the information needed to recover. This adds a 120-second default for Step run_command, cleans up discoverable Linux descendants on timeout or abort, reports signal termination as failure, and preserves a readable full log when command output is shortened.

The change also keeps runtime state from interfering with solution delivery:

• STEP_CODING_AGENT_PLAN_DIR optionally stores plans outside the workspace. Existing defaults and saved session paths remain authoritative.
• Compaction preserves prior history and refuses empty replacement summaries. Summary output respects the model limit and the existing 32,000-token ceiling, including when a larger reserve is used to trigger compaction earlier.
• --completion-check git-committed checks delivery in the same session. Optional --completion-review shares the existing follow-up budget.
• File-exploration guidance recognizes the active Step aliases search_files, find_files, and list_directory. This removes shell-only fallback instructions when dedicated tools are available, while preserving guidance for actual shell-only profiles.

Explicit command timeouts, background execution, and permission decisions keep their contracts. Linux descendant snapshots cannot atomically contain already-reparented descendants or forks racing the snapshot. This change adds no dependency.

Validation: the updated branch includes main at 93ebc5b, retaining its bounded summary-overflow recovery, loaded-skill preservation, and final tool-output handling. The compaction merge keeps the 32,000-token per-request ceiling, prior split-turn history, and independent empty-history/prefix rejection. The resolved file received an independent source review; npm run check and the complete offline ./test.sh suite pass on the merged tree. The suite builds workspace outputs offline and uses an isolated provider-free environment with a three-CPU affinity limit.

The prompt correction has 65 passing targeted tests, including six alias cases first reproduced as failing on the prior implementation. Legacy, mixed, shell-only, and PowerShell profiles remain covered. The lockfile and generated model data are unchanged.

Benchmark results apply to their pinned runtime versions; the appended prompt correction has not been evaluated in a new full benchmark run. This PR makes no causal score-improvement claim. Denial-recovery configuration is tracked separately in #188.