fix: resolve 5 critical architectural issues (race condition, name-based PDF fill, AppError handler, OLLAMA_MODEL env var)

[Gopisokk]

Summary

Resolves 5 architectural issues identified by auditing the full codebase.

---

Changes

fix(filler) — Replace fragile index-based PDF filling with name-based matching

• Previously: annot.V = answers_list[i] — matched purely by position/index
• Now: reads annot.T (the PDF field label), decodes the PDF string, looks up value directly in the LLM answers dict (case-insensitive)
• None → writes "" (not the string "None"), list values → joined with "; ", unmatched fields → warning log instead of silent wrong fill

fix(file_manipulator) — Eliminate race condition from shared mutable LLM state

• Previously: one LLM() instance in self.llm was mutated per-request — a data corruption race condition under concurrent requests
• Now: a fresh LLM(transcript_text=..., target_fields=...) is created inside fill_form(), scoped to each request
• Also converts the silent return None on missing PDF to a proper FileNotFoundError

fix(forms) — Remove duplicate get_template() DB call

• get_template() was called twice per request — doubled DB round-trips for every fill
• Now fetched once, stored, reused
• FileNotFoundError is caught and surfaced as a 422 response

fix(main) — Register AppError exception handler

• register_exception_handlers(app) was defined in api/errors/handlers.py but never called in api/main.py
• Every AppError raise (e.g. 404 Template not found) produced an unhandled 500 crash instead of a proper JSON error

fix(llm) — Make Ollama model configurable via OLLAMA_MODEL env var

• Model name was hardcoded to "mistral" — operators had to edit source code to change it
• Now reads os.getenv("OLLAMA_MODEL", "mistral") — backward compatible, defaults to "mistral"
• Moved host/model resolution outside the per-field loop (was being re-evaluated on every iteration)

---

Relates to: #44