mail-negotiator: classify merchant replies, draft gated, audited sends

- IMAP poll (UNSEEN, stdlib imaplib) -> parse (bare address via email.utils)
- threading: In-Reply-To -> outbound sent-draft message-id -> proposal
- classify: deterministic rules; optional LLM behind provider interface;
  below-threshold confidence -> needs_review (never guessed)
- unthreadable approvals flagged needs_review (suspicious by definition)
- merchant_status recorded on proposal detail; proposal lifecycle untouched
- reply drafts generated from templates (drafted), NEVER auto-sent
- SMTP sends only draft_status=approved (human approve_draft, audited);
  sent drafts store their Message-ID for future threading
- idempotent on Message-ID; DaemonRun wrapper for monitor

tests: 9 new (classification rules, threshold->review, threading, idempotence,
blocked send without approval, approve->send->audit chain, double approve)
— 29 total passing

cayceops/components/maild/classify.py

@@ -0,0 +1,68 @@
+"""maild classification: rules first, LLM optional, conservative threshold.
+
+The LLM never decides alone: any result below ``CAYCE_OPS_LLM_REVIEW_THRESHOLD``
+confidence becomes ``needs_review``. With the ``null`` provider (default) only
+deterministic rules run.
+"""
+
+from __future__ import annotations
+
+import json
+import re
+
+from cayceops.config import Config
+from cayceops.models import MailClassification
+
+_RULES: list[tuple[MailClassification, re.Pattern[str], float]] = [
+    # (class, pattern, confidence when matched)
+    (MailClassification.REJECTION, re.compile(r"\b(declin|reject|not able to|unable to|unfortunately|we regret|won.t be able)\w*", re.I), 0.9),
+    (MailClassification.APPROVAL, re.compile(r"\b(approv|accept|welcome to the|you.re (in|approved)|live|activated|glad to)\w*", re.I), 0.85),
+    (MailClassification.QUESTION, re.compile(r"\b(could you|can you|please (provide|clarify|confirm)|what (is|are)|how (do|does|can)|need more (info|details))\b", re.I), 0.7),
+]
+
+_LLM_PROMPT = """Classify this merchant email into exactly one of:
+approval, rejection, question, other.
+Reply with strict JSON: {"classification": "...", "confidence": 0.0-1.0}
+
+Subject and body:
+{body}"""
+
+
+def _rules(text: str) -> tuple[MailClassification, float]:
+    for kind, pattern, confidence in _RULES:
+        if pattern.search(text):
+            return kind, confidence
+    return MailClassification.OTHER, 0.3
+
+
+def _llm_classify(text: str, cfg: Config) -> tuple[MailClassification, float] | None:
+    """Ask the configured LLM. Returns None on any failure (rules take over)."""
+    from cayceops.llm import get_provider
+
+    try:
+        provider = get_provider(cfg)
+        raw = provider.complete(_LLM_PROMPT.format(body=text[:4000]), system="You classify merchant emails. Reply with JSON only.")
+        data = json.loads(re.search(r"\{.*\}", raw, re.S).group(0))
+        kind = MailClassification(data["classification"].lower())
+        confidence = float(data["confidence"])
+    except Exception:  # noqa: BLE001 — LLM is best-effort by design
+        return None
+    if kind not in MailClassification or kind is MailClassification.NEEDS_REVIEW:
+        return None
+    return kind, min(max(confidence, 0.0), 1.0)
+
+
+def classify(text: str, cfg: Config) -> tuple[MailClassification, float]:
+    """Classify merchant text. Degrades to rules when no LLM is configured."""
+    if cfg.llm_provider != "null":
+        llm_result = _llm_classify(text, cfg)
+        if llm_result is not None:
+            kind, confidence = llm_result
+        else:  # LLM failed — fall through to rules
+            kind, confidence = _rules(text)
+    else:
+        kind, confidence = _rules(text)
+
+    if confidence < cfg.llm_review_threshold:
+        return MailClassification.NEEDS_REVIEW, confidence
+    return kind, confidence