Plan written to .claude/plan.md. Here's the summary:

The core success pattern learning engine (sw-success-patterns.sh, 586 lines) and test suite (26 tests) are already built. Four targeted gaps need closing:

1. Acceptance tracking is dead code — sw-pipeline.sh:2894 greps for recommended_template: in the state file, but nothing writes it. Need to persist the recommended template after intake.

2. Cost always 0 — success_capture_pattern() hardcodes cost_usd: 0. Need to read the pipeline's computed cost via env var.

3. No success correlation — Tracks acceptance/rejection but not whether accepted recommendations led to successful pipelines. Need success_track_correlation() function + recommendations_succeeded counter.

4. No issue_type field — Patterns capture labels/complexity but not explicit issue type (bug/feature/etc). Need to extract from intake-metadata.json.

Approach: Minimal targeted fixes (~80 lines across 3 files). All changes are additive JSON fields with jq // 0 defaults — fully backwards compatible, no migration needed.

10 tasks, ending with test verification. The plan includes alternatives considered (embedding-based matching rejected as overkill), risk analysis, data flow diagram, and rollback strategy.