Pipeline B (Template-Driven Extraction) - Implementation Tracker

[abhishek-8081]

Tracker for implementing Pipeline B (template-driven extraction) on the
development-approach-b branch, following the architecture decided in discussion #609.

Approach B, in short

No standalone extraction step. The transcript (from text input #546 / voice input #547) is
stored as the source of truth. Each template holds its own field dictionary (field name →
type). At form-fill time, the stored transcript + the template's field dictionary go to the
LLM, which returns values field-by-field, which are then drawn into the PDF. The LLM runs
per form at fill time. No canonical incident schema, no incidents layer, no reporting layer
(those are Pipeline C, on the approach-c branch).

Context

Most of Pipeline B already exists as working, tested prototype code (per-field LLM loop,
PDF filling, template registry, sync + async fill endpoints). The work here is mainly
reshaping that prototype into the v1 API layering (routes → services → repositories) and
wiring it to the stored transcript, plus giving B its own contract.

Sub-issues

• [Pipeline B] Reshape templates & forms OpenAPI contract to Approach B #637 Reshape templates & forms contract to Approach B
• [Pipeline B] Wire the fill flow to accept input_id and use the stored transcript #638 Wire the fill flow to accept input_id and use the stored transcript
• [Pipeline B] Link FormSubmission to Input by foreign key #639 Link FormSubmission to Input by foreign key
• [Pipeline B] Apply routes → services → repositories layering to forms & templates #640 Apply routes → services → repositories layering to forms & templates
• [Pipeline B] Route Controller/FileManipulator through the form service #641 Route Controller/FileManipulator through the form service
• [Pipeline B] Fix positional PDF field-matching in Filler #642 Fix positional PDF field-matching in Filler