Motir

The foundation every other story routes through: until a lesson can be keyed to the card being planned, nothing can leave the rules corpus — moving a paragraph out would only make it unreachable.

The existing category axis is dead data, proven on origin/main: selectForInjection never passes categories, captureMistake writes categories: [], and the only values in existence are hand-written topical tags on the seed rows that nothing queries. A topical tag also answers the wrong question — what a lesson is about, never which card it applies to.

So key the lesson the way the rule packs already route. MANIFEST.md's selector is pack(operation, phase, kind, type); the two corpora should share one routing model rather than invent a second:

The phase axis is the one with no predecessor and it discriminates hardest. Mistakes made while laying a level's children are a different family from mistakes made while writing a card's body: the first are about shape, edges and coverage; the second about criteria, sizing and claims. Today a planner writing a description receives all the widening lessons.

All three are set-valued, empty means unconstrained — the packs' own routing is set-valued (type ∈ {code, chore, deploy}), so a migration lesson carries three types and a scalar would force it to "applies to everything".

⚠️ phase is overloaded. PlannerPhase already means onboarding_planning | regular_planning — which pass is running, not which part of the card is being written. mistakeTypes keeps that job as the consumer pre-filter; the three axes sit underneath it.

Acceptance criteria

• Lesson carries kinds, types and phases, set-valued, defaulting to empty; empty on an axis means unconstrained on it, and the three are independent.
• Each vocabulary matches the work item's; phases is exactly skeleton / deepen. A test pins all three so a typo fails the build rather than never matching.
• listForInjection applies one overlap clause per axis, each omitted when its axis is empty; selectForInjection carries all three and passes them down.
• mistakeTypes behaviour is unchanged, asserted.
• A test proves the axes narrow independently, that a lesson tagged for three types matches a query for one of them, that a skeleton query returns widening lessons and not deepening ones, and that an unconstrained lesson reaches every query.
• Existing rows are tagged by a forward-only, idempotent data migration; the dead categories column's disposition is recorded rather than silently repurposed.

Context refs

• motir-ai src/services/lessonService.ts — selectForInjection (no categories), captureMistake (categories: []), PlannerPhase.
• motir-ai src/repositories/lessonRepository.ts — the && overlap clause with omit-when-empty, the pattern the three axes follow.
• motir-ai prisma/schema.prisma — the Lesson model.
• motir-meta prompts/plan-rules/MANIFEST.md — pack(operation, phase, kind, type).
• motir-ai fly.toml release_command · .github/workflows/ci.yml:107.