[engine] mode chooses the review: "llm" (default, unchanged), "classifier" (Jev alone, no LLM)
or "combined" (Jev, then the LLM with Jev's flagged units as hints in the user prompt).
werkrevisor splits each changed PR document into ##/### sections and #### Scenario units,
asks Jev once per unit with the operator's questions, and maps every answer to a problem value:
each question asks for a problem, a score runs from "no problem" to "worst problem".
A clear spec problem is RED, a doubtful one or an oversized unit YELLOW, a Jev error ERROR.
The hints carry file, line, scenario id, question and value, but no PR text.
tools/jev-eval measures the questions offline on local PR documents.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
