[arch-migration] Umbrella: WWDC-gated phases (P2–P5, P7) — re-assess after 2026-06-08 keynote

[roryford]

Umbrella for the WWDC-gated phases of the target-architecture migration. These are deliberately not broken into detailed issues yet — Apple's 2026-06-08 keynote (agent/tool/on-device-media APIs) may reprioritize the driver and media seams, so they get assessed next week and split into detailed issues then.

Plan: docs/plans/target-architecture-migration.md (target in docs/plans/target-architecture.md). Docs PR: #1604.

Gated phases

• P2 — Engine carve. Done via downward Contract extraction (kept ManifoldInference name; no ManifoldEngine module). See docs/plans/p2-engine-carve-split.md.
  • [arch-migration P2a] Extract ManifoldContract leaf module (thin Contract kernel via downward extraction) #1719 — P2a: extract ManifoldContract leaf (thin Contract kernel) ✓ merged
  • [arch-migration P2b] Grow P0c characterization harness (canonicalizer fields + handoff/usage goldens) #1720 — P2b: grow P0c harness (canonicalizer + handoff/usage goldens) ✓ merged
  • [arch-migration P2c] De-tangle ConversationTurnExecutor (the real refactor) #1721 — P2c: de-tangle ConversationTurnExecutor — seams extracted ✓ merged (refactor(ManifoldRuntime): de-tangle ConversationTurnExecutor into per-turn seams #1724)
  • [arch-migration P2c-followup] Extract TurnCompressionCoordinator + thin flows (grow compression goldens first) #1725 — P2c-followup: TurnCompressionCoordinator + thin flows (compression goldens first)
• P3 — Turn-Driver seam + resumable ConversationRun. TurnDriver + SingleTurnDriver (behaviour-preserving) + ResumableRunDriver + RunStore port + schema bump. Multi-agent/plan-execute drivers remain deferred.
• P4 — Modality generify. Generic MediaGeneration<Output> + collapse MessagePart generated-media cases (schema migration). Parallelizable.
• P5 — Trait → product. Retire MCP/Voice/Tools/AppIntents cleanly; extract Ollama/CloudSaaS source first. Lockstep edits to PackageTraitGateAuditTest + FeatureMatrixTests + build-modes.yml.
• P7 — Retire @_exported shims (final breaking release, after ≥2-minor deprecation window).

Action next week

After the keynote, re-confirm sequencing and split each phase into detailed issues (these are substantial, not minor changes).

[roryford]

Post-WWDC re-assessment (June 9, 2026)

WWDC keynote was June 8. Here is the sequencing re-assessment for the gated phases.

What changed

Foundation Models: AFM 3 is multimodal (image input). The image API is not yet visible in the current installed SDK (Xcode 26.4 pre-WWDC) — a post-WWDC beta likely ships it. Tracked separately by #1710.

Core AI: Confirmed iOS 27/macOS 27. Does not unblock any of P2–P7 today. Tracked by #1577.

Metal 4 / MLX: M5 Neural Accelerator requires macOS 26.2+. Tracked by #1711. No architectural implications for P2–P7.

Swift 6.2 SE-0470: Isolated conformances — available now. Tracked by #1712. Informs P2 (Engine carve) refactor shape.

Phase re-assessment

Phase	Status	Notes
P2 — Engine carve	Still blocked on P0c harness (#1605 prereq)	SE-0470 (#1712) can clean up conformances; consider doing that first as a low-risk warm-up
P3 — Turn-Driver + resumable run	Not yet unblocked	Deferred pending P2
P4 — Modality generify	Partially unblocked	AFM 3 multimodal (#1710) is the first new modality; P4 can absorb it once P0c is stable
P5 — Trait → product	Not yet unblocked	No WWDC changes to trait surface
P7 — Retire shims	Not yet	Unchanged

Recommended near-term sequencing

1. refactor(ManifoldInference): adopt SE-0470 isolated conformances for backend protocols #1712 (SE-0470, unblocked, low risk) — before Engine carve to reduce scope
2. feat(ManifoldMLX): expose M5 Neural Accelerator availability in capability probe #1711 (MLX accelerator gate, unblocked)
3. feat(ManifoldFoundation): wire AFM 3 image-input API (WWDC 2026) #1710 (Foundation multimodal, blocked on SDK) — once post-WWDC Xcode beta drops
4. P2 Engine carve — when P0c harness is solid

Splitting P2 into detailed sub-issues is deferred until P0c characterization harness (#1566 area) is stable. The carve is the one real refactor here and persona-review is mandatory.

[roryford]

P2 split + path decision (June 9, 2026)

Post persona plan review (concurrency / module-graph / behavior-preservation lenses), P2 is split into three separate issues: #1719 (P2a), #1720 (P2b), #1721 (P2c). Full writeup: docs/plans/p2-engine-carve-split.md.

Key reframe — thin Contract by extracting down, not evicting the engine up. The review measured the engine-up move (the original P2 framing) at 246 lockstep files / 35 modules with cycle risk, because InferenceService is welded to the kernel by BackendRegistrar + concrete use from MCP/Llama/MLX. Extracting the Contract into a new leaf ManifoldContract (the proven P1 pattern from #1608/#1611) reaches the same dependency shape at ~30 files / 0 consumer import-edits.

Naming decided: keep ManifoldInference — no rename to ManifoldEngine, no module to fold. The clean 3-layer shape (ManifoldContract ← ManifoldInference ← ManifoldRuntime) falls out of the downward extraction alone, which dropped the original MED-risk rename/rehome phase.

Sequencing: (#1719 ∥ #1720) → #1721. The first two are parallel-safe (disjoint files; serialize pushes); #1721 is sequential (needs the grown goldens, rewrites the turn loop, persona-reviewed diff).