Issue: Add schema-bound LLM classifier for Gmail candidates

Feature Description

Add an internal classifier boundary that lets a Gmail domain extractor ask an LLM whether a bounded candidate message or thread is relevant to the user's request. The classifier must return strict JSON and include grounded evidence.

Motivation

Deterministic scoring is useful for candidate pruning, but the LLM should make semantic relevance decisions. This is required for ambiguous requests such as "emails related to closing" where keyword matching alone confuses real estate closings with account closing, deal closing, or unrelated business language.

Proposed Solution

• Define a GmailCandidateClassification JSON schema.
• Add an internal LLM call boundary that does not recursively start a full agent loop.
• Classify candidates after metadata ranking and before ledger writes.
• Validate model output against source text.
• Preserve weak matches separately from accepted records.
• Add benchmark fixtures for travel and real estate false positives.

Acceptance Criteria

• Classifier returns JSON with is_relevant, confidence, reason, evidence, entities, dates, amounts, identifiers, and next_actions.
• Non-JSON model output is retried or rejected.
• Evidence must be grounded in candidate content.
• Local-model benchmark reports precision, recall, latency, refusal rate, and token usage where available.