feat: complete v1 — platform-aware scaffold + MNIST/CIFAR-10 demos

[SamPlvs]

Summary

• v1 is complete. All 8 PRD §9 acceptance criteria met, all 6 STATE.md Known Issues closed.
• Platform-aware Docker scaffold — scaffold_delivery(gpu_enabled=...) auto-detects host GPU and emits the correct compose template (CPU on macOS / no-GPU Linux, GPU otherwise). Fixes the broken deploy.resources.devices: capabilities: [gpu] block on Mac.
• Two v1 demo deliverables — MNIST 99.66% test accuracy (Tier 3, 64s on MPS) and CIFAR-10 91.62% (Tier 3, 427s on MPS). Full Phase 1-6 deliverables in separate delivery repos.

What changed

Platform — src/zo/scaffold.py + src/zo/cli.py

Split _COMPOSE_TEMPLATE into GPU and CPU variants. New _resolve_compose_template() helper picks based on gpu_enabled: bool | None. CLI's _init_commit_writes probes via detect_environment() and passes through. Service name kept as gpu across both for cross-platform README parity.

Tests — tests/unit/test_scaffold.py + tests/unit/test_cli.py

6 new TestPlatformAwareCompose tests (explicit modes, auto-detect both ways, detection-failure fallback, CPU service-name parity). One existing test made host-independent.

Demos — plans/cifar10-classifier.md + delivery repos (separate)

CIFAR-10 plan added (public benchmark, no client data). Both demo delivery repos live at ~/Documents/code/personal/{mnist-digit-classifier,cifar10-classifier}-delivery/ — outside this repo, not part of this PR. Each has full Phase 1-6 artifacts (reports/{data_quality,training,analysis,model_card,validation}.md), trained weights in two formats × two locations (PyTorch + ONNX, redundancy against corruption), drift detection scaffold, and oracle + unit + export tests.

Memory

• STATE.md — Known Issues feat(platform): Phase 2 + 3 — memory, semantic index, orchestration engine, lifecycle wrapper #3, feat(platform): Phase 4 — evolution engine, CLI, integration tests #5, docs: README rewrite + MNIST toy project for e2e validation #6 closed; v1_status: COMPLETE.
• DECISION_LOG.md — 4 new decisions (platform-aware scaffold, stale-issue cleanup, v1 milestone, demo deliverables).
• PRIORS.md — PR-033 (platform-aware templates rule) + PR-034 (MPS-pytest tensor extraction quirk).
• New sessions/session-023-2026-04-26.md per CLAUDE.md "On Session End" protocol.

Docs

README test count 669 → 675, demos badge added, status section refreshed.

Verification

Check	Result
ZO platform pytest	675 passed, 7 skipped
MNIST delivery pytest	16/16 passed
CIFAR-10 delivery pytest	19/19 passed
ruff check src/zo/	clean
validate-docs.sh	10/10, 0 failures
14 ZO CLI commands smoke-tested	all working
zo preflight on both plans	6/7 PASS each (only nvidia-smi WARN)
PyTorch + ONNX exports load + round-trip	SUCCESS for both demos
zo watch-training Rich Live dashboard	renders against ZOTrainingCallback output

v1 acceptance criteria (PRD §9)

• First project's oracle criteria fully satisfied — MNIST 99.66%
• Delivery repo contains zero ZO infrastructure (only .zo/ portable state)
• DECISION_LOG provides complete audit trail
• Session recovery works
• Self-evolving mechanism triggered ≥ 1 time — 34 priors documented
• All 11+ project delivery agent definitions written (20 total)
• All 6 platform build agent definitions written
• Plan.md edit during execution succeeds

Test plan

• PYTHONPATH=src pytest tests/ -q — 675 passed
• ruff check src/zo/ — clean
• bash scripts/validate-docs.sh — 10/10
• Train MNIST end-to-end → 99.66%, 16 tests pass
• Train CIFAR-10 end-to-end → 91.62%, 19 tests pass
• Smoke-test all 14 ZO CLI commands
• Verify PyTorch weights save + load + forward pass for both demos
• Verify ONNX onnx.checker.check_model for both
• Verify zo watch-training Rich Live dashboard renders
• Confirm scaffold emits CPU compose on Mac (no broken capabilities: [gpu])

🤖 Generated with Claude Code