canvas-humanizer-loop

# canvas-humanizer-loop Wrap `canvas-humanizer` in a bounded repair loop. The purpose is to catch damage introduced by rewriting, not to keep rewriting until an external score changes. ## Inputs and boundaries Accept the same paths and voice inputs as `canvas-humanizer`, plus: - `course_id`, `assignment_id`, and the stable assignment `work_dir`; - optional `spec_path`, `rubric_path`, `source_paths`, and `hard_locks`; - `max_iter`, default 3 and hard-capped at 3. Require `work_dir` to end in `course-<course_id>__assignment-<assignment_id>`. Write loop artifacts below `<output_dir>/_loop_iter_<N>/` and `loop_log.json` beside the final output. Validate the path against `src.course_artifacts.stable_work_dir` before writing. This skill performs local file work only. It does not call Canvas, write an assignment `result.json`, or declare submission readiness. The caller must run its course-level verification after the loop. Use `src.humanizer_segmentation.split_sentences` and `paragraph_segment_counts`; do not invent another splitter. ## Iteration layout ```text _loop_iter_1/ humanized.docx humanizer_log.json audit-a.json audit-b.json audit-c.json merged-audit.json repair-log.json _loop_iter_2/ ... loop_log.json ``` Use atomic JSON/text writes. Keep the original draft and every iteration so the best safe artifact can be recovered. ## Stage 1: preflight 1. Read the complete original, spec, rubric, and necessary source excerpts. 2. Confirm hard locks are uniq...