feat(notes): consolidate note generation on a shared spec and improve structure

[Valtora]

Description

A pass over Nojoin's note-generation prompts to improve the consistency and quality of generated meeting notes.

The core problem: the automatic first pass (the unified meeting-intelligence prompt) and the "Regenerate Notes" button (the standalone notes prompt) used two separate, drifted prompt bodies, so regenerating notes could produce a different structure from the original. This change makes both paths share a single source of truth and adopts an evidence-based note structure.

What changed:

• Added NOTES_BODY_SPEC in backend/utils/meeting_notes.py as the single source of truth for note structure and fidelity rules. Both note-generation prompts (unified meeting_intelligence.py and standalone base.py) embed it verbatim, so they can no longer drift. A drift-guard test asserts both contain it.
• Restructured notes to a best-practice, follow-through-oriented layout: Summary first, a dedicated scannable Key Decisions section (with rationale), prominent Action Items (single named owner plus due date), then per-topic Detailed Notes, then Miscellaneous. This replaces the previous topic-list-first layout that buried decisions.
• Added anti-hallucination fidelity rules to the notes body (never invent facts, decisions, figures, or attributions; attribute claims to the speaker who made them; favour signal over volume). The primary path previously had none.
• Notes now begin at the Summary section and no longer emit their own top-level title heading, which duplicated and sometimes contradicted the meeting title the app already displays. Enforced by the prompt instruction, the JSON schema example, a defensive strip_leading_title_heading() helper applied in both paths, and an updated result contract (notes must start with a ## section heading).
• Unified the persona to "expert meeting-notes assistant" across the notes and title prompts, and routed the standalone title prompt through the shared short-title instruction so both title paths behave consistently.
• Raised the Anthropic standalone-notes max_tokens from 2048 to 4096 to match the unified path so comprehensive notes are not silently truncated.
• Removed the dead legacy speaker-table prompt and its helpers (get_default_speaker_prompt_template, get_speaker_prompt_template, build_speaker_prompt, parse_mapping_table); the evidence-JSON speaker-suggestion path fully supersedes them and they had no production or test callers.

No new dependencies.

Type of change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that changes existing behaviour)
• Documentation update

Checks run

• Backend tests: source .venv/bin/activate && pytest (864 passed)
• Python quality: python scripts/check.py (Ruff lint, format check, mypy, doc and Alembic validators)
• Frontend lint: cd frontend && npm run lint (not run: no frontend paths changed)
• Frontend unit tests: cd frontend && npm run test (not run: no frontend paths changed)
• Frontend build: cd frontend && npm run build (not run: no frontend paths changed)
• Docs validation: python3 scripts/validate_docs.py
• Alembic validation: python3 scripts/validate_alembic.py

Migration impact

• No database migration in this PR.

Documentation impact

• No documentation change required. Confirmed no guide specifies the note section layout; docs/USAGE.md describes only what the notes-language setting controls (still accurate) and docs/ARCHITECTURE.md covers prompt caching (unaffected).

Security impact

• No security-sensitive change.

Manual verification

Automated coverage only in this PR (864 backend tests pass, including new drift-guard and title-strip tests). Recommended live smoke check before or after merge: process a recording and use "Regenerate Notes", confirming the notes open at the Summary heading, carry a scannable Key Decisions section, and contain no duplicate title line. No capture, context-menu, auth, or migration paths are touched.