Learn: post-campaign retrospective that writes deduped, durable lessons back into briefs and strategy

[wesleysimplicio]

Parent epic: #64

Goal

Port simplicio-learn's retrospective pattern to campaigns: after each campaign window (or piece batch), mine high-signal lessons from journals, gate results, and analytics, dedupe them against existing memory, and write them where the next run actually reads them — so lessons are applied, not re-derived.

Sources

• Piece journals (Journal: durable run-journal + stall detector so the loop switches strategy instead of oscillating #66): which strategy switches worked, which gates fail most per channel/format.
• Analytics scores (Analytics: score organic performance across social, Reddit, forums, and dev portals #54) and winner/loser classification (lib/promotion/classifier.ts).
• Compliance/watcher-gate violations: recurring violation types per client.
• Community replies (Community loop: monitor comments and draft respectful technical replies #60): recurring objections and FAQs.

Destinations (durable, lean)

• .marketing-engine/clients/<slug>/learnings.md — client-specific bullets (voice, claims that passed/failed compliance, best hooks per channel).
• data/learnings.jsonl — structured entries feeding brief generation and the strategy ladder ordering in Journal: durable run-journal + stall detector so the loop switches strategy instead of oscillating #66.
• Channel registry notes (Channels: build social + developer community registry for autonomous publishing #49) — per-channel etiquette lessons from rejections.

Rules

• Durable, reusable bullets only — no transcripts, no one-off anecdotes.
• Every lesson carries evidence refs (journal/analytics paths) and a confidence tag; UNVERIFIED lessons never override MEASURED ones (claims discipline from lib/gate/claims-gate.ts).
• Dedupe before write: an equivalent existing lesson gets its evidence count bumped, not duplicated.

Acceptance criteria

• E2E test: a mocked campaign run produces ≥1 lesson with evidence refs; running the retrospective twice does not duplicate it.
• Brief/piece generation demonstrably consumes learnings (e.g. a "losing hook pattern" lesson excludes that pattern in the next batch — asserted in test).
• Memory stays lean: retrospective enforces a max size per learnings file, compacting lowest-confidence entries first.