feat(email): quality & connection metrics logging

[itomek]

Log and export quality and connection diagnostics that feed the evaluation harness.

Acceptance criteria

• Logs and can export: categorization results, false positives, false negatives, and connector/connection diagnostics.
• Metrics feed the evaluation harness.

Test acceptance criteria

• CI asserts false-positive rate <5% and false-negative rate <2% against the feat(eval): generate + commit labelled email-triage corpus + gemma4 baseline #1230 corpus; the eval harness fails when either is exceeded.

Consumed by the LLM-judge (#1279) and CI (#1112).

[github-actions]

This is in scope and most of the plumbing already exists — this issue is mainly about adding an export sink + ground-truth comparison, not building logging from scratch.

What's already there to build on:

• Categorization results — src/gaia/agents/email/verbose.py already structured-logs every decision via log_triage_decision(...) with category/confidence under the gaia.agents.email logger (extra={"stage": ...}). The taxonomy + group_by_category() live in src/gaia/agents/email/tools/triage_heuristics.py.
• Connection diagnostics — src/gaia/agents/email/gmail_backend.py surfaces token/scope/connector state through gaia.connectors (see the revoked-scope path around line 415).
• Eval feed — src/gaia/eval/scorecard.py already aggregates pass/fail by category (by_category), so wiring exported records into it is the natural seam.

Suggested shape:

1. Add an export path that serializes the existing verbose records to JSON/CSV (one choke-point, mirroring verbose.py's design).
2. Compute false positives/negatives by diffing auto-categorizations against labeled fixtures — this is the "Triage category accuracy" success metric in docs/plans/email-triage-agent.mdx §23, and the eval harness scenarios in §21.3.

Related: #1108 (email triage scenario category) is the eval-harness counterpart — worth coordinating so the export format matches what the scenario loader expects.

A maintainer should confirm whether the labeled-fixture corpus (§21.3's 200-message fixture) is the intended ground-truth source here.