RFAI-03: Proposal generator V1 with verified provenance and outcomes ledger

[Chris0Jeky]

Context

Week 3 from taskdeck-12-week-roadmap-v4.md.

Parent: #972
Depends on: #974

This issue turns envelopes into typed proposal batches and makes provenance verifiable instead of decorative.

Scope

• Add IProposalGenerator returning typed ProposalBatch output with field-level provenance.
• Distinguish extractive fields from inferred fields.
• Verify extractive quote/span claims against the originating source block with fuzzy matching.
• Verify inferred fields by resolving EvidenceLink references emitted for the envelope.
• Wire Microsoft.Recognizers.Text as a deterministic pre-extraction layer for dates, numbers, durations, URLs, emails, and similar constraints.
• Persist proposal provenance in the proposal path.
• Add proposal_outcomes persistence for approval/edit/reject/ignore decision events.
• Measure deterministic pre-extraction impact; do not claim round-trip reduction without data.

Acceptance Criteria

• Proposal batches include provenance for every generated field.
• Extractive provenance is rejected or downgraded when quote/span verification fails.
• Inferred provenance is rejected or downgraded when evidence links do not resolve.
• Date hallucination regression cases from the golden set pass.
• proposal_outcomes records every proposal action with content-free/bucketed dimensions.
• Golden-dataset smoke reports schema validity, extractive quote-match rate, and inferred evidence-resolution rate separately.

Suggested Verification

• Backend unit tests for provenance verifier branches
• API/application tests for proposal persistence and outcomes ledger
• Golden dataset smoke command introduced by RFAI-01: Safety invariants, IA cut, eval seed, and recertification #973/RFAI-02: IntentEnvelopeV1, IChatClient adapter, and schema spike #974