Task: Encode the per-component codegen drift-elimination method as a skill

A reusable skill that runs one component through the full drift-elimination algorithm. The skill is the single method every remediation story follows, so the sweep does not re-derive the recipe per component and does not repeat the pilot mistakes. It encodes the algorithm the ores.refdata and ores.dq pilots proved (see the parent story's Decisions, 2026-09-03), the generated-code review step, the too-complex-to-generate exception rule, and the automation ramp with its evidence gates.

• The skill is authored under doc/llm/skills/<slug>/SKILL.org per the claude_code_skills.org contract and deployed to .claude/skills/ via deploy_skills. It loads through the Skill tool.
• The algorithm is stated as an ordered procedure: baseline regen before any model edit, survey every model file, bind entities to profiles, record each no-match with its real feature values, fix codegen gaps at their source, regenerate, require byte-identical output, review the generated code for C++ idiomaticity and smells, extend the check_component_drift.py gate list last. No step is skippable; no partial coverage is a valid end state.
• The too-complex-to-generate exception rule is stated: an exception needs a recorded justification per entity, and the manual artefact stays visible to the drift check.
• The automation ramp is stated as levels L0-L2 with the evidence each level needs before promotion.
• The skill cites its provenance: the ores.refdata and ores.dq worked method records, and the sprint 24 dimension settlement.
• The dormant product-backlog entries that this method supersedes (entity_evaluation_skill and similar) are closed or re-pointed at the skill.

• 2026-09-03: task scaffolded with the rollout task (see its Plan).
• Name the skill against the naming registers: no register verb covers repair to a defined end state, so extend the verb register with fix-, recording the decision in skill_naming_conventions.org's decision log (per the register's own extension mechanism).
• Scaffold via compass add skill under doc/llm/skills/<slug>/ and author the whole method from the provenance records only: the refdata and dq worked records, the sprint 24 dimension settlement, and the story's 2026-09-03 Decisions.
• Wire: catalogue regeneration, s1-agent cross-link, dormant entity_evaluation_skill backlog entry re-pointed at the skill.
• Verify with deploy_skills (export + id-link resolution) and a site export check before raising.

Manual QA scenarios (scaffolded via compass add test_scenario, run through the QA Validation Runner panel) that verify this task. Link new ones here as they're created; the scenario doc itself links back via its "Verifies task" field.

Closed 2026-09-04. The canonical method ships as the compass-codegen-fix-drift skill (doc/llm/skills/compass-codegen-fix-drift/SKILL.org, ID 7908E2AF-E196-443F-8FB7-0047553D8EFE) in this PR (c5f51aeb9f, 62cbd692d7, 400620bbf0, a4e2bd8fc5). The name extends the skill naming verb register with fix-, recorded in that document's decision log. The content distills the refdata and dq worked records, the sprint 24 dimension settlement, and the story's 2026-09-03 Decisions into seven ordered, unskippable steps: baseline regen before any model edit, survey every model file, compute resolved feature values, bind every match and record every no-match with its real values, scripted per-entity regeneration with zero-additional-diff and idempotency requirements, generated-code review for C++ idiomaticity (template fixes only), and the drift gate extended last. It states the too-complex-to-generate exception rule (recorded justification per entity, artefact stays visible to the drift check) and the L0-L2 automation ramp with its evidence gates.

Acceptance: authored under doc/llm/skills/<slug>/SKILL.org per the skill contract and deployed via deploy_skills (clean export to .claude/skills, id-links resolve); ordered procedure with no skippable steps and no partial-coverage end state; exception rule; L0-L2 ramp with evidence gates; provenance citations (refdata and dq worked records, sprint 24 settlement); the dormant entity_evaluation_skill backlog entry re-pointed at the skill — all met. The skill is also wired into the catalogue and the s1 agent.