Restores two pieces of vanilla behavior the React port silently dropped:
1. The output divs were contenteditable in vanilla — clinicians fix
AI mistakes inline before saving the encounter or copying out.
The early React port rendered the output as a read-only div, so
any edit forced a copy-paste workflow.
2. correctionTracker.js (public/js/correctionTracker.js) saved every
meaningful inline edit to user_memories under category
correction_<section> so the AI learns user preferences. Settings
→ Corrections still showed them, but nothing in React was *writing*
them — the loop was broken.
New EditableResult component wraps the result body in a textarea,
captures the AI baseline on first render / after refine|shorten, and
on blur posts to /api/memories/correction (only when the edit clears
the noise threshold from vanilla — wordDiff ≥ 2 OR charDiff ≥ 20 on
outputs > 100 chars).
Wired into all 7 note pages:
Encounter → section: 'encounter'
Dictation → section: 'hpi'
SOAP → section: 'soap'
SickVisit → section: 'sickvisit'
WellVisit → section: 'wellvisit'
HospitalCourse → section: null (server enum has no correction_hospital)
ChartReview → section: null (server enum has no correction_chart)
Tests:
shared/clinical/correction-tracker.ts (+ .test.ts) — pure heuristic
with vectors covering the noise floor (single-word swap on long
text → skipped; new sentence → tracked; large char diff → tracked).
Lives in shared/ so the root vitest config picks it up; the React
component imports from there.
e6d0e5ef8c