Add roadmap plan for long-running pipeline logging

[Copilot]

This change documents a concrete plan for improving observability in long-running graphify operations. It focuses on the current gaps in pipeline and watch-mode progress reporting so future implementation work has clear scope and success criteria.

• Roadmap addition

  • Added a new ROADMAP.md entry: 3.4a Long-Running Process Observability & Logging
  • Positions the work next to incremental/watch-mode improvements, where the logging need is most visible

• What the plan covers

  • stage-level timing for detect/extract/build/cluster/analyze/export
  • heartbeat-style progress updates during long extraction loops
  • throughput signals such as files/sec and extraction volume
  • watch-mode diagnostics for debounce, re-detect, merge, cluster, and export phases
  • warning/failure summaries that stay concise by default and expand under --verbose

• Implementation shape

  • outlines a lightweight timing approach around existing stage boundaries
  • proposes a shared progress-reporting abstraction for PipelineRunner and WatchMode
  • keeps the design compatible with a later move to ILogger / telemetry-backed diagnostics

• Rollout guidance

  • defines a phased rollout:
    1. add elapsed-time summaries
    2. add periodic heartbeats
    3. centralize structured diagnostic fields

Example target output shape from the documented plan:

[2/6] Extracting code structure...
      412/1204 files processed (34%), 18 skipped, 27.4 files/sec, elapsed 00:00:15

[2b/6] Running AI-enhanced semantic extraction...
      21/87 files processed, 3 warnings, elapsed 00:01:42

[watch] 6 file(s) changed
        debounce: 500ms | extract: 1.2s | merge+cluster: 340ms | export: 220ms