Preserve deep + flat eval runs and add a comparison toggle

[sadacca]

Previously the output filename was {article_id}{version}{model_sig}.json with no mode component, so a deep run and a flat run of the same article+models wrote to the same path — the second silently overwrote the first, making side-by-side comparison impossible.

• bias_eval: encode mode in the output filename ({article_id}{version}{model_sig}_{mode}.json) so deep and flat results coexist on disk. Non-destructive: _already_evaluated's glob and build_index's content-based reads are unaffected; legacy files keep their names.

• article_eval page:

  • Add a deep / flat / both mode selector. "both" evaluates each mode in sequence in the background thread, writing one file per mode, and only runs modes not already cached.
  • Replace the single-result loader with _load_results_by_mode(), which discovers every stored run for the article and keys them by mode (legacy untagged files treated as deep).
  • Completion polling now waits for all requested modes, with an "N/M modes complete" progress hint.
  • Results view gains a mode toggle (shown only when >1 mode exists), a caption showing the displayed run's mode + models, and mode-suffixed download names.

Claude-Session: https://claude.ai/code/session_01669jmknuhhZY3vBqwpDWvF