Contrôle qualité des messages de prospection

# Outreach QA & scoring gate

You have the "lead-scorer" MCP server connected (Lead Scorer CRM, authenticated with Lead Scorer OAuth at the endpoint configured by the installed plugin). Use its tools for CRM reads and writes. For public web research, use your host assistant's native web search and browsing, then save verified findings with the CRM tools. Discover resource IDs with the available list/search tools; never guess or probe sequential IDs, and ask me when no discovery tool exists. Never invent data: if a tool result is empty, say so. An API key is only a manual fallback for clients without OAuth support.

## Goal
Nothing in campaign <CAMPAIGN_ID> reaches me for approval until it has been scored and, if needed, rewritten. Grading your own drafts is not optional: it is the step that separates outreach from spam.

## The rubric (100 points)
| Axis | Points | What earns them |
| --- | --- | --- |
| Grounded personalization | 30 | A dated, sourced signal appears in the first two sentences and drives the argument |
| Swap test | 20 | Pasting another lead's name breaks the message |
| Single ask | 15 | Exactly one CTA, phrased as a question, low friction |
| Length & rhythm | 15 | Under 150 words (email) / 600 characters (LinkedIn); mixed sentence lengths |
| Honesty | 10 | No invented number, customer, or shared history; no unearned claim |
| Voice | 10 | No corporate speak, no AI tells, sounds like one person writing to another |

## Steps
1. Resolve the campaign with `list_campaigns` when needed, then call `get_campaign` and `list_campaign_actions` to pull every draft.
2. `get_campaign_authoring_context`: you need the source signals to judge whether the personalization is real or hallucinated. **A message that references a signal not present in the context scores 0 on Honesty and is flagged, not fixed silently.**
3. Score every draft, axis by axis. Show the table.
4. **Rewrite everything under 70** with `update_campaign_action_draft`, then re-score. Two rewrites maximum: a third failure means the lead lacks a real signal, so flag it for removal instead.
5. **Campaign verdict:**
   - Median ≥ 80 and no honesty flags → ready for my approval
   - Median 70-79 → ship, but name the weakest axis so the next campaign fixes it upstream
   - Median < 70, or any honesty flag → do not present for approval; the problem is research, not copy
6. **Report.** Score distribution, worst axis, the 3 lowest drafts before/after, and any lead you recommend removing.

## Hard rules
- Never raise a score because a rewrite "feels better". The rubric decides.
- Never silently delete a flagged draft. Flagging is the deliverable.