refactor(triage): agent-supplied category + exclusion + canonical context + owner-input rule

[br4vesirrobin]

Summary

Pivots the triage feature (from #7) to an agent-in-the-loop model and adds precision controls so that being from an important sender is not enough — a task is created only when the email genuinely requires the owner's input.

Built by deep-interview intent → implement → code-review (APPROVE) → live-verified against a real mailbox + Asana.

What changed

• Agent is the classifier. Removed the in-process LLM categorizer (judge_category/httpx/llm.*) from triage. The agent reads candidates and supplies the CategoryDecision. classifier.py + llm.* are untouched (they still power folder classification).
• Two-command interface (the pipeline auto-hook is removed — headless process can't classify):
  • kontor-cli triage — lists qualifying candidates (with bodies + canonical folder), read-only.
  • kontor-cli triage-create --email-id --category [--deadline] [--no-dry-run] — creates one idempotent Asana task (dedup marker kontor-id:<message-id|uid> scoped to the target project). Defaults to dry-run.
• Exclusion rule — is_automated() heuristic + config triage.exclude_senders (no-reply/notifications/digests/*.mail.microsoft/Office365Reports/Google/Miro/Azure/…). Notification & platform mail never becomes a task.
• Customer rule tightened — a bare external sender no longer auto-qualifies; it needs content signal. The deterministic gate is explicitly a recall net.
• Owner-input precision rule — triage.owner_email + SKILL rubric: create a task only when the content requires the owner's personal input/action (decision/review/approval/answer). Sender importance alone is never sufficient — this is the agent's judgment.
• Canonical-structure context — each candidate carries its taxonomy folder (e.g. 2_Projects/PRJ_Willemen) from the deterministic rules engine (lazy, best-effort) to inform categorization.
• Cleanup — triage_enabled now gates the CLI; removed the vestigial triage.scan_rebuild.

Live verification (real mailbox → Asana)

• Gate after exclusion tuning: 140 → 79 candidates (automated noise dropped; real customers like Willemen/Eiffage kept).
• Precision rule proven: 5 "important sender but no action" emails (a "thanks" reply, four "approved" confirmations) → 0 tasks; 5 genuine asks (PR review on a customer blocker, budget approval, AI investment input, license approvals) → 5 correctly-categorized tasks. Re-run → skipped_dedup.

Tests

228 tests pass, coverage 94% (≥90 gate), ruff check + ruff format --check + mypy --strict clean. Code-reviewer: APPROVE (2 LOW findings fixed in the follow-up commit).

🤖 Generated with Claude Code

[br4vesirrobin]

Autobahn forward-port completed against current main at 68c47bc.

Key repairs in the refreshed head 1e5f9b8:

• replaced the deleted legacy rules-engine integration with the current canonical Pipeline.classify_with_rules() path;
• preserved agent-supplied categorization and dry-run-by-default behavior;
• made real Asana writes validate configured projects before mailbox access;
• made Asana dedup/create API failures fail nonzero instead of being reported as skipped;
• enforced exact dashed YYYY-MM-DD deadlines before mailbox access;
• added current AI-SDLC intake, specification, handoff, traceability, and workflow evidence;
• expanded regression coverage, including the --no-dry-run write path.

Local verification on the published head:

• 249 tests passed
• Ruff check and format check passed
• mypy passed
• AI-SDLC validation, rules validation, and archgate passed
• independent architecture and code-review passes approved the exact head

Fresh hosted checks are now running; this PR will be merged only if they are all green.