style-migrate-enhance

name	style-migrate-enhance
type	agent-invocable
description	High-throughput migration waves converting priority parent CSL 1.0 styles to Citum with repeatable before/after metrics and migration-engine gap recommendations. Fidelity is the hard gate.
model	sonnet

Style Migrate+Enhance

Authoritative shared process docs:

• docs/policies/STYLE_WORKFLOW_DECISION_RULES.md
• docs/guides/STYLE_WORKFLOW_EXECUTION.md
• docs/architecture/MIGRATION_STRATEGY_ANALYSIS.md — converter role and limits

Use This Skill When

• The task is portfolio migration (long-tail / dependent styles).
• You need repeatable before/after/rerun metrics.
• You want concrete recommendations for citum_migrate improvements from observed gaps.
• You need to produce a migrate seed for a subsequent tune pass on an embedded-core style.

Role of Migrate Output

For dependent styles, the migrate output is the deliverable (subject to the fidelity hard gate). For embedded-core styles, the migrate output is a seed and evidence source feeding a subsequent style-tune pass — it is not a finished embedded style. See docs/architecture/MIGRATION_STRATEGY_ANALYSIS.md for the strategic basis: the converter is an evidence tool, not the canonical authoring path for high-impact styles.

Input Contract

• Legacy style path(s) under styles-legacy/.
• Target Citum style path(s) under styles/.
• Batch size and priority source.
• Optional target metric.

Output Contract

• Updated style YAML file(s).
• Shared metrics and rerun evidence in the format described by the shared execution guide.
• Migration-pattern gaps and recommended converter/preset follow-up when observed. Every CSL 1.0.2 type reaches styles as its real ref_type under the conversion contract (docs/specs/CSL_TYPE_CONVERSION_CONTRACT.md), so when the source CSL style handles a type like collection, review, performance, or figure, a missing type-variants entry in the migrated YAML is a reportable migration gap, not converter noise.

Autonomous Operation

Run the full wave without pausing between styles. Use the shared docs for the common evidence order, decision rules, and output shape. Only interrupt for Cargo.toml/Cargo.lock changes or git push origin main.

Workflow

1. Select the next priority wave.
2. Seed the baseline with the smallest trustworthy evidence surface.
3. Apply the fix according to the shared policy and execution guide.
4. Re-run apples-to-apples comparison evidence.
5. Treat supplemental rich-input evidence as confirmation when configured.
6. Commit each passing style and produce final metrics plus follow-up recommendations.

Hard Gates

• Never accept a fidelity regression.
• Never classify a registered divergence as a migration or engine bug without updating adjudication first.
• SQI is tie-breaker and optimization only for dependent styles. For embedded-core targets, SQI is not a gate here — the tune pass owns SQI to green.
• After bounded retries with no progress, note it in the wave summary and move to the next style rather than halting the entire wave.

Required Artifacts

• Iteration log.
• Final wave summary table.
• Code Opportunities table in the same shape as the router skill when engine gaps are observed.

Verification

• Structured oracle: node scripts/oracle.js <legacy-style> --json
• Core quality report: node scripts/report-core.js
• Supplemental official style report for configured rich-input styles: node scripts/report-core.js --style <name>
• Optional full workflow impact: ./scripts/workflow-test.sh <legacy-style>