canvas-humanizer-surgical

# canvas-humanizer-surgical Apply a targeted second pass. Do not reopen the whole draft when only a small set of segments is damaged. ## Inputs and boundaries Require: - `draft_path`, `output_path`, and prior `humanizer_log.json`; - `course_id`, `assignment_id`, and stable assignment `work_dir`; - voice register and relevant spec/rubric/source anchors; - explicit issue list or an audit artifact identifying target segments; - optional hard locks and banned-phrase list. Require `work_dir` to end in `course-<course_id>__assignment-<assignment_id>`. Keep a source copy and write `surgical_log.json` beside the new output. Validate the path against `src.course_artifacts.stable_work_dir` before writing. This is local file transformation only. Never call Canvas, mutate an external platform, write assignment `result.json`, or claim the assignment is ready. Use the canonical sentence utilities from `src.humanizer_segmentation`, lock logic from `src.humanizer_segment_extract.extract_locks`, and deterministic scoring from `src.humanizer_score`. ## Stage 1: audit and role tagging Map every target to `(doc_paragraph_index, intra_para_index)` and verify its exact current text. Reject stale issue records whose anchor no longer matches. Classify target roles conservatively: - `intro_opener`, `intro_thesis`, `intro_setup`; - `body_topic_sentence`, `body_quote_lead_in`, `body_quote_follow_up`; - `body_elaboration`, `body_anecdote`, `body_closing`; - `conclusion_thesis_restate`, `concl...