dsh-observation-journal: Leaving Running Facts for DSH Sessions

dsh-observation-journal is a pure observer plugin for DeepSeek Harness (DSH), designed to record runtime facts of agent tasks. This plugin registers no tools, invokes no LLM calls, and does not intervene in agent decision-making; it only writes the task title, model hierarchy, tool calls, failure counts, duration, and final status to a human-readable journal file after the session concludes. Core features include: maintaining auto-generated statistical blocks with markers in the journal while preserving manually edited regions; generating an append-only JSONL raw sidecar file containing the complete planning trajectory, tool counts, and task hashes for subsequent LLM-driven insight analysis. The plugin incorporates cross-process write locking and stale lock reclamation mechanisms to ensure write reliability.