Import Training Plan — Resume from Audit-Flagged Partial Ship to Full Phase 1 Ship in 14 Hours

The import-training-plan feature originally shipped 2026-04-16 as a partial-ship: the parser, mapper, orchestrator, and 23 unit tests landed cleanly, but the source-picker and preview views were never wired into navigation and confirmImport() was a no-op stub. Audit UI-015 (2026-04-20) caught it. Resume attempt 2026-05-06 then surfaced a deeper structural error: the original PRD claimed ImportOrchestrator writes to TrainingProgramData — structurally impossible because TrainingProgramData is a struct with only static let fields. The honest path was a mid-flight rollback to research, a rewritten PRD against the actual persistence target (EncryptedDataStore.importedTrainingPlans), and a re-decomposed task list.

The same session also produced a v4.X skill-layer upgrade as a meta-byproduct. A user-ordered pre-Phase-4 audit caught 4 P0 spec errors that would have hit "no such symbol" at compile time (AppRadius.pill, AppMotion.standardEase, SettingsActionLabel with custom badge slot, toast component — all referenced by the spec, none in the codebase). The user requested promoting that audit pattern to a mechanical PM-workflow gate. The result: /ux preflight + /design preflight + /ux pre-merge-review + /design pre-merge-review + /design build auto-dispatch, all shipped in PR #235 and adopted by this feature on first run.

4 PRs landed across 2 repos in a single ~14-hour session · 18/44 tasks done · 33 new tests · 4 Figma frames auto-built (the v4.X auto-dispatch flow's first production output) · 0 P0 in make ui-audit · 4 P0 spec errors caught before code was written.

Architecture decisions (locked, not re-litigated)

1. Persistence target = EncryptedDataStore.importedTrainingPlans — sixth @Published collection in the encrypted store, persisted via the existing 2-phase commit pattern. Closes the structural PRD gap.
2. Routing layer = TrainingProgramStore — gains activePlanId: UUID?. exercises(for:in:) checks the flag; nil returns bundled program, non-nil returns the active imported plan's exercises. Mutual exclusion enforced.
3. Domain model = ImportedTrainingPlan (Identifiable, Codable) — separate from the parser-transient ImportedPlan. Wraps [ImportedDayAssignment] (each carries the user-editable assignedDayType).
4. GDPR coverage — Article 17 (delete) free transitively via EncryptedDataStore.deletePersistedData() extension. Article 20 (export) via 3 touch points in DataExportService.swift. Sync semantics deferred to Phase 2.

What shipped

Outcomes (tier-tagged)

Five honest disclosures

1. The original PRD's persistence claim was structurally impossible. v1 PRD said "ImportOrchestrator writes to TrainingProgramData" — but TrainingProgramData is a struct with only static let fields. No write path. The audit caught the partial ship; the resume caught the architectural gap. The fix was a full mid-flight rollback to research → rewritten PRD → re-decomposed tasks.

2. The user-ordered pre-Phase-4 audit caught 4 P0 spec errors before any code was written. Audit cost ~20 min; Phase 4 rework cost would have been ~2-4h. This pattern is now mechanical via /ux preflight + /design preflight (shipped in PR #235).

3. The /design build Figma auto-dispatch was the v4.X chain's first production run. The flow worked end-to-end on first invocation: preflight → MCP liveness check → page creation → 4 mobile-screen frames → node ID write-back → state.json + figma-code-sync-status.md updates → PR description gate satisfaction. One iteration was needed (clipping fix on Frame 3).

4. Three patterns net-new to the codebase. .swipeActions, .contextMenu, and a bespoke ImportedPlanRow component were never used in FitMe before this feature. They're now documented in feature-memory.md as design-system evolutions.

5. Phase 2 sync is deferred. CloudKit and Supabase per-record sync for imported plans is out of Phase 1 scope. The needsSync: Bool field exists on the model from day one so Phase 2 can opt records in without a schema migration.

Lessons for future features

1. PRDs that name persistence targets must reference the actual write path (file:line), not just the type name. v1 said "writes to TrainingProgramData"; v2 says "writes to EncryptedDataStore.importedTrainingPlans via the existing 2-phase commit pattern, touch points: EncryptionService.swift:779-1062 (5 locations)". The latter is verifiable by Phase 6 review; the former isn't.
2. Pre-flight existence checks should be mechanical, not manual. The /ux preflight + /design preflight gates added in v4.X catch this class of error before Phase 4 begins.
3. Figma sync is part of Phase 3, not a "deferred follow-up". Auto-dispatching /design build makes Figma sync part of the contract.
4. Honest scope rejection beats expanding ambition mid-resume. The 1-day rollback cost was net positive.
5. deferred task status is honest, not failure. Phase 2 sync is deferred with rationale and tracking — better than faking Phase 1 done with a stub.

Cross-cutting framework signals

• v7.5 / v7.6 / v7.7 / v7.8 — All write-time gates fired correctly during the resume. Zero gate skirts.
• v4.X Skill-Layer Upgrade — Shipped as a meta-byproduct. Promoted 4 audit patterns to mechanical gates. Phase 3 chain extended 7→11 steps; Phase 6 chain 4→5 steps. Phase 7 BLOCKED unless both pre-merge reviews pass.

Where things go from here

• Phase 2 follow-up PRD (out of Phase 1 scope) — CloudKit per-record sync + Supabase table + per-day editor + AI prompt regeneration + PDF/photo/share-extension sources.
• First post-launch metrics review scheduled for 2026-05-13 (T+7d) — query GA4 for import_started → import_completed → import_plan_activated funnel; compute activation rate; verify kill criteria thresholds aren't tripped.
• Backfill figma_node_ids for already-shipped features is now a normal part of the framework and will happen as features get touched.

Closing

This feature is two case studies in one. The surface case study is the persistence + active-plan + GDPR architecture closing audit UI-015. The deeper one is how the framework caught a structural PRD error mid-flight, demanded honest rework, and emerged with a v4.X skill-layer upgrade that promotes 4 audit patterns to mechanical gates so the next feature doesn't relearn the same lesson. The trigger event (4 P0 spec errors) became the test case for the new gates within the same session.

Full source case study with timeline detail, hard/easy retrospective, and per-task notes: docs/case-studies/import-training-plan-case-study.md.