fix: map refinement sequence to target for response routing

Fixes P1 code review issue where refinement transcriptions were being
discarded due to missing sequence→element mapping.

When uploading combined audio for Phase 2 refinement, we now properly
map the returned sequence number to the target element via
`context.transcriptionTargets[sequenceNum]`. This allows the refinement
response handler to correctly identify and update the target field.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: rosscado

src/state-machines/DictationMachine.ts

@@ -1466,6 +1466,7 @@ const machine = createMachine<DictationContext, DictationEvent, DictationTypesta
           Object.keys(targetTranscriptions).forEach(key => {
             const seq = parseInt(key, 10);
             delete context.transcriptions[seq];
+            delete context.transcriptionTargets[seq]; // Clean up target mapping too
           });
 
           // Store the refined transcription
@@ -1977,6 +1978,9 @@ const machine = createMachine<DictationContext, DictationEvent, DictationTypesta
             `[DictationMachine] Refinement transcription sent with sequence ${sequenceNum} for target ${targetId}`
           );
 
+          // Map the refinement sequence to the target element so the response can be routed correctly
+          context.transcriptionTargets[sequenceNum] = targetElement;
+
           // Track this as an active refinement
           context.activeRefinementSequences.set(targetId, sequenceNum);