Agent drain robustness

[CodingInAVan]

Discovery scans one grouping level deep, so a standalone agent watching a parent folder finds both flat / and grouped run-// layouts.
Keep one bad session from stalling the launcher's 60s drain cap:

• LogTailer finishes a session whose .tmp/ working dir has been frozen past a stale grace (producer crashed or was killed), not only when it is removed.
• TailerManager settles each finished session with a marker discovery skips:
  .uploaded (clean) or .failed (orphaned). GPUFL_AGENT_PRUNE_FAILED deletes an orphaned session instead - its windows are already uploaded.
• A launcher-spawned --upload agent (GPUFL_AGENT_IGNORE_PREEXISTING) uploads only sessions newer than its JVM start, ignoring old sessions in a reused output dir.