ADR-0035: --fix Mastery Repair and Partial Pillar Eligibility

Status

Accepted (2026-10-22)

Context

palee validate --fix recomputes topic_mastery via applyMasteryRepairs, which zero-fills absent pillars before calling computeTopicMastery. The valid-topic-mastery rule does the same. This means a note with 3 of 4 pillars scored (e.g. conceptual/practical/debug: 0.9, no feynman) computes (0.9 + 0.9 + 0.9 + 0) / 5 = 0.54 and can be flagged or rewritten as drift — locking a learner out of a gate on the strength of a pillar they never entered.

The issue (#514) asks: does an absent pillar mean "zero in that dimension" or "no data"? ADR-0021 already resolved this for adopt and review: absent pillars are omitted, not zeroed, because 0 is a score and "absent" is not.

Decision

Absent pillars mean "no data", not zero. This direction is adopted from #277 (carriedPillarScores omits absent pillars) and #300 (record-review drops them). --fix and valid-topic-mastery must agree: a topic with any absent pillar is skipped, not zero-filled. The rule and repair only operate on topics where all four pillars are present and shaped.

This is a "skip" decision, not a "reweight" decision: reweighting the formula (c + p + d + 2·f) / 5 to sum over a subset of pillars would change the weighting semantics and silently alter mastery for notes whose formula was always meant to run over all four. Skipping is conservative and matches the rule's existing behavior for the all-absent case.

Consequences

• Rule (valid-topic-mastery.ts): changed from "skip if all pillars absent" to "skip if any pillar absent." Topics with partial data are not flagged for mastery drift.
• Repair (validate.ts applyMasteryRepairs): unchanged in mechanics; it only repairs topics the rule flagged, so partial-pillar topics are naturally excluded. The : 0 fallback in the pillar map is now unreachable but retained as a type guard.
• assessed_at is never touched by --fix — unchanged. The timestamp attests to an assess recording, not a validate recomputation.
• Tests: updated the two "Greptile P1" tests that asserted zero-fill behavior; added a 3-note fixture test (#514 reproduction) covering no-pillars, partial-pillars, and full-pillars in one validate --fix run.
• See also: ADR-0021 (assessment surface), #277, #300, #514.