Assessment Pipeline: Add L1 (Eliminatory) filters, post-processing & results view #187

Describe the current behavior
The assessment flow only supported a single LLM pass (prompt + config → results).
There was no way to:

• Pre-filter rows before the expensive LLM batch (e.g. drop off-topic or duplicate
  submissions), so every row hit the model regardless of relevance.
• Post-process results (derive scores, filter, sort) without exporting to a
  spreadsheet and doing it by hand.
• View results comfortably — the preview opened in a cramped modal.

For high-volume datasets this means wasted LLM cost on irrelevant rows and manual
spreadsheet work after every run.

Describe the enhancement you'd like

• Add an Eliminatory (L1) pre-filter stage before evaluation:
  • Topic Relevance — gate that excludes REJECT rows from evaluation.
  • Duplicate Detection — passthrough check reported in the export.
• Add a Post Processing stage (computed columns, filter, sort) usable both
  before the run and on completed runs.
• Open results in a dedicated full-page spreadsheet view instead of a modal.
• Surface L1 pass/reject stats and the latest model per saved config.