feat(voice): write a note only when the clinician asks for one

The comment instruction told the model to sweep up leftovers: "clinical
notes, in the language spoken. Omit the parts already captured as treatment
type, teeth, prosthesis work or deadline." So filler, small talk and an
unrequested diagnosis all became a persisted clinical note — and comment is
the one durable trace a recording leaves (§10).

The model now reports the exact words that asked for the note in a new
commentTrigger field, and comment holds only what was dictated after them.
resolveVoiceIntent keeps comment only when a trigger is present, so the rule
is enforced by code instead of trusted to the prompt: a model that decides
leftover speech was a note loses that note.

Per-locale trigger vocabulary goes in the prompt's locale notes, beside the
tooth vocabulary — «بنویس که», "write this in the notes", "noteer". The
resolver only checks that a trigger was reported, so it needs no per-locale
knowledge and stays locale-neutral, as §6 requires.

Also adds HARD RULE 7: never decide something is a note; speech that fits no
field is simply not reported.

Four resolver tests. The two that drop an unasked note fail without the
guard; the trigger-only and trigger-with-note cases pass either way and exist
to prove the guard does not swallow a real note.

Frontend untouched — it only ever saw the resolved comment, never the
trigger.

Spec: §5 gains the rule and the field, decision 54 added.

Gates: backend 220 tests, nest build, ESLint clean on the voice module;
frontend tsc --noEmit clean, 52 Vitest tests.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

backend/src/modules/voice/extraction.prompt.ts

@@ -10,16 +10,22 @@ const LOCALE_NOTES: Record<string, string> = {
     '"شش بالا راست" = upper right six -> arch "upper", side "patient_right", position 6.',
     'A jaw is spoken as "فک بالا" (upper jaw) or "فک پایین" (lower jaw), sometimes just',
     '"بالا"/"پایین" in context, or "هر دو فک" (both jaws).',
+    'A note is asked for with "بنویس", "یادداشت کن", "این را یادداشت کن", "بنویس که",',
+    '"در توضیحات بنویس". Anything else the clinician says is not a note.',
   ].join(' '),
   nl: [
     'The clinician is speaking Dutch, where FDI is standard. "zesentwintig" and "26" are',
     'tooth 26. The descriptive form is "rechtsboven zes" = upper right six. A jaw is',
     '"bovenkaak" (upper) or "onderkaak" (lower), or "beide kaken" (both).',
+    'A note is asked for with "schrijf op", "noteer", "zet in de notities". Anything else',
+    'the clinician says is not a note.',
   ].join(' '),
   en: [
     'The clinician is speaking English and uses FDI. "twenty-six", "two six" and "26" are',
     'all tooth 26. The descriptive form is "upper right six". A jaw is "upper jaw"/"lower',
     'jaw", or "both jaws".',
+    'A note is asked for with "write this in the notes", "note that", "add a note",',
+    '"put in the comments". Anything else the clinician says is not a note.',
   ].join(' '),
 };
 
@@ -110,6 +116,9 @@ export function buildExtractionPrompt(
     '   one is not.',
     '6. A whole jaw is not a tooth. Never put a jaw in the "teeth" list. A jaw appears only',
     '   as a prosthesis[].targets entry with "arch" set and "position" null.',
+    '7. Never decide something is a note. Fill "comment" only when the clinician asked for',
+    '   one, and report the words they used in "commentTrigger". Speech you cannot place in',
+    '   a field is simply not reported — it is never a note.',
     '',
     'TOOTH NUMBERS',
     'A number the clinician says for a tooth IS that tooth\'s FDI code. Put it in "fdi" as',
@@ -160,8 +169,13 @@ export function buildExtractionPrompt(
     '',
     'OTHER FIELDS',
     '- connectedSpans: only for bridges or splinted units. Endpoints inclusive.',
-    '- comment: clinical notes, in the language spoken. Omit the parts already captured as',
-    '  treatment type, teeth, prosthesis work or deadline.',
+    '- comment + commentTrigger: a note is written ONLY when the clinician asked for one.',
+    '  "commentTrigger" is the exact words that asked, copied from the transcript. "comment"',
+    '  is what they then dictated, WITHOUT those words: for "بنویس که بیمار حساسیت به سرما',
+    '  دارد", commentTrigger is "بنویس که" and comment is "بیمار حساسیت به سرما دارد".',
+    '  If nobody asked, BOTH are null. Never sweep up leftover speech, filler, small talk or',
+    '  a diagnosis nobody asked you to record. A comment without a trigger is discarded, so',
+    '  guessing costs the clinician the note.',
     '- labMatchExact: true only when the spoken name matched a lab name exactly.',
   ].join('\n');
 

backend/src/modules/voice/extraction.resolver.spec.ts

@@ -505,6 +505,7 @@ describe('resolveVoiceIntent', () => {
     teeth: [tooth('14'), tooth('15')],
     connectedSpans: [],
     comment: '  حساسیت به سرما  ',
+    commentTrigger: 'بنویس که',
     prosthesis: [],
     labId: null,
     labMatchExact: false,
@@ -679,6 +680,40 @@ describe('resolveVoiceIntent', () => {
     );
   });
 
+  it('drops a note the clinician never asked for', () => {
+    // The whole point: leftover speech the model decided was a note is discarded. Before this
+    // rule the prompt told it to sweep up "the parts already captured" as a comment.
+    const result = resolveVoiceIntent(
+      { ...base, comment: 'بیمار عصبی بود', commentTrigger: null },
+      CTX,
+    );
+    expect(result.comment).toBeNull();
+  });
+
+  it('drops a note whose trigger is only whitespace', () => {
+    const result = resolveVoiceIntent(
+      { ...base, comment: 'بیمار عصبی بود', commentTrigger: '   ' },
+      CTX,
+    );
+    expect(result.comment).toBeNull();
+  });
+
+  it('keeps a note that was asked for, trimmed', () => {
+    const result = resolveVoiceIntent(
+      { ...base, comment: '  حساسیت به سرما  ', commentTrigger: 'بنویس که' },
+      CTX,
+    );
+    expect(result.comment).toBe('حساسیت به سرما');
+  });
+
+  it('reports no note when the trigger was heard but nothing followed it', () => {
+    const result = resolveVoiceIntent(
+      { ...base, comment: '   ', commentTrigger: 'بنویس که' },
+      CTX,
+    );
+    expect(result.comment).toBeNull();
+  });
+
   it('resolves a due date through the same context', () => {
     const result = resolveVoiceIntent(
       { ...base, due: { kind: 'weekday', weekday: 'thursday', which: 'this' } },

backend/src/modules/voice/extraction.resolver.ts

@@ -545,8 +545,15 @@ export function resolveVoiceIntent(
   const due = resolveDueDate(intent?.due, ctx.todayIso, ctx.weekStartJs);
   if (due.unresolved) unresolved.push(due.unresolved);
 
+  // A note is written only when the clinician asked for one. The model reports the words that
+  // asked ("بنویس که", "write this in the notes"); without them, whatever it put in `comment`
+  // is leftover speech it decided was a note, and it is dropped. Matching the phrase itself
+  // stays in the prompt, so this check needs no per-locale vocabulary.
+  const commentAsked =
+    typeof intent?.commentTrigger === 'string' &&
+    intent.commentTrigger.trim().length > 0;
   const comment =
-    typeof intent?.comment === 'string' && intent.comment.trim()
+    commentAsked && typeof intent?.comment === 'string' && intent.comment.trim()
       ? intent.comment.trim()
       : null;
 

backend/src/modules/voice/extraction.wire.spec.ts

@@ -23,6 +23,7 @@ const wire = (overrides: Partial<WireVoiceIntent> = {}): WireVoiceIntent => ({
   teeth: [],
   connectedSpans: [],
   comment: null,
+  commentTrigger: null,
   prosthesis: [],
   labId: null,
   labMatchExact: false,

backend/src/modules/voice/extraction.wire.ts

@@ -50,6 +50,7 @@ export type WireVoiceIntent = {
   teeth: WireToothIntent[];
   connectedSpans: { from: WireToothIntent; to: WireToothIntent }[];
   comment: string | null;
+  commentTrigger: string | null;
   prosthesis: WireProsthesisAssignment[];
   labId: string | null;
   labMatchExact: boolean;
@@ -125,6 +126,7 @@ export const VOICE_INTENT_JSON_SCHEMA = {
     'teeth',
     'connectedSpans',
     'comment',
+    'commentTrigger',
     'prosthesis',
     'labId',
     'labMatchExact',
@@ -154,7 +156,17 @@ export const VOICE_INTENT_JSON_SCHEMA = {
     },
     comment: {
       type: ['string', 'null'],
-      description: 'Clinical notes, in the spoken language.',
+      description:
+        'The note the clinician explicitly dictated, in the spoken language, WITHOUT the ' +
+        'words that asked for it. Null unless they actually asked for a note. Never put ' +
+        'leftover speech here.',
+    },
+    commentTrigger: {
+      type: ['string', 'null'],
+      description:
+        'The exact words that asked for a note, copied from the transcript (e.g. ' +
+        '"بنویس که", "write this in the notes"). Null when nobody asked. A comment with ' +
+        'no trigger is discarded.',
     },
     prosthesis: {
       type: 'array',
@@ -298,6 +310,7 @@ export function toVoiceIntent(wire: WireVoiceIntent): VoiceIntent {
     teeth: teeth.map(toToothIntent),
     connectedSpans,
     comment: wire?.comment ?? null,
+    commentTrigger: wire?.commentTrigger ?? null,
     prosthesis: assignments.map(toProsthesisAssignment),
     labId: wire?.labId ?? null,
     labMatchExact: wire?.labMatchExact === true,

backend/src/modules/voice/voice.types.ts

@@ -66,6 +66,12 @@ export type VoiceIntent = {
   teeth: ToothIntent[];
   connectedSpans: ConnectedSpanIntent[];
   comment: string | null;
+  /**
+   * The exact spoken words that asked for a note — "بنویس که", "write this in the notes".
+   * Null when nothing asked. The resolver keeps `comment` only when this is present, so the
+   * model cannot decide on its own that leftover speech was a note.
+   */
+  commentTrigger: string | null;
   /** Empty array, never null. */
   prosthesis: ProsthesisAssignment[];
   /** Must be one of the linked-lab ids supplied in the prompt, or null. */