Motir

Move the retrieval question from the pass down to the card, on all four coordinates.

selectForInjection takes queryText and its own comment calls it "the planning job's intent" — one string for the whole pass, so twelve cards planned in a run embed the same query twelve times and get identical lessons. This card replaces that with a per-card question:

• queryText built from the card — title, kind, type, target repo, the deliverables its description enumerates.
• kinds / types from the card being planned.
• phases from what the pass is doing: skeleton while laying a level's children, deepen while writing a body. This is the coordinate only the pass knows and it has never been asked for.

Know what this half can and cannot do. The ranker returns INJECTION_LIMIT_DEFAULT = 8 ordered by cosine distance, so the query text does most of the selecting — and a harness composing text from card fields can only ever write in feature register ("board filter at scale"), while lessons are stored as takeaways ("a count taken from a working tree is not a property of the ref"). Those embed far apart.

So this card's job is a sound baseline, not the best possible retrieval: correct axes, a per-card query, no regression. The register translation — describing the risk rather than the feature — needs judgement the harness does not have, and belongs to the planner's own search (MOTIR-3321). Say so in the code rather than leaving a later reader to conclude the query builder is underdone.

Keep the composition in one shared helper beside selectForInjection; both planner call sites go through it today and the comment says why. Preserve the two properties the design leans on: selected ids stay recorded per card on the job record, and an empty selection still renders nothing.

Acceptance criteria

• One exported helper builds the per-card query and the three axes; generate, augment and replan all call it and none constructs its own.
• The phase passed reflects what the pass is doing — skeleton when laying children, deepen when writing a body — asserted for both.
• selectForInjection is called once per card rather than once per pass.
• Selected lesson ids are recorded per card on the job record.
• A card with no type still produces a usable query and does not constrain that axis.
• The helper's own comment states that its query is feature-register and that risk-register search is the planner's tool, so the limitation is recorded rather than read as an oversight.
• An empty selection yields an empty promptSection and injects nothing.
• pnpm typecheck clean; the planner-input suites pass.

Context refs

• motir-ai src/jobs/plannerInputs.ts — where selectForInjection is called and queryText is filled today.
• motir-ai src/repositories/lessonRepository.ts — INJECTION_LIMIT_DEFAULT = 8 and the cosine ORDER BY that make query quality decisive.
• motir-ai src/services/lessonService.ts — SelectForInjectionInput, the one-shared-helper contract.
• motir-ai src/jobs/handlers/generateTree.ts, src/jobs/handlers/replan.ts — the call sites.
• MOTIR-3300 — the axes this passes · MOTIR-3321 — the pull half that supplies the register this one cannot.