feat(storage): add Phase 5 SQLite contract preview

[Yuge Zhang (ultmaster)]

Note

Reworked onto current main after #93 merged, so this diff contains Phase 5 only.

What changed

• Add an isolated node:sqlite implementation of the structured storage contracts.
• Cover lifecycle, strict schema v1, transactional migrations, Space/Node/log/Task repositories, atomic ordered writes, deletion fencing, and connection-lifetime anti-resurrection.
• Add the missing reusable Tasks/Runs contract and align Disk and SQLite with the Run-completion semantics now present on main.
• Reuse refactor(storage): Phase 4.5 — move storage-owned layout inside the boundary #93's backend-neutral naming owner and move Canvas record-shape validation to a backend-neutral owner.
• Recognize SQLite as an available preview adapter while deliberately keeping it non-selectable.
• Update the storage architecture and authoritative roadmap for the merged Phase 4.5 baseline and completed Phase 5 preview.

Reading the new backend

Nine implementation modules land under backends/sqlite/, plus contract and integration tests, test support, and the immutable v1 migration fixture:

File	Owns
structured-store.ts	StructuredStore — adapter lifecycle and composition
space-repository.ts	SpaceRepository
space-nodes.ts	SpaceNodes
space-tasks.ts	SpaceTasks
space-logs.ts	SpaceEvents, SpaceChanges, and the delta log
space-write.ts	SpaceHandle.write
database.ts	connection, schema v1, migrations, deletion fencing, and tombstones
rows.ts	row statements and JSON codecs
identity.ts	title/label collision-key allocation

The first six align with the port surface and reuse the Disk filenames for the same roles, so the parity suites in ports/contracts/ map directly across both adapters. identity.ts remains pure allocation logic and imports #93's backend-neutral naming rules, keeping collision behavior identical without importing the Disk adapter.

Why

This provides a concrete contract and churn preview before SQLite is wired into runtime composition, without changing product capabilities. It tests whether the extracted storage seam supports a transactional backend before committing to production selectability.

Selecting sqlite still fails in validateStorageProfile, now with a distinct preview diagnostic: the adapter is available but not selectable. Physical Disk reads, World bootstrap, Blob placement, import/export, and Workspace remounting still require later composition work.

Scope

• 25 files, +3,879 / −247 against current main
• Confined to apps/server/src/modules, docs/architecture/canvas-storage.md, and docs/proposals/multi-backend-storage.md
• No web/shared protocol, UI, runtime composition, Disk-to-SQLite migration/import, Postgres, Azure, Agenetes, or RFS changes
• Disk remains the only selectable structured backend

Validation

All re-run on the published head (2bc7a6bb):

• pnpm run check — passed
  • server: 120 files, 1,096 passed / 2 todo
  • web: 131 files, 990 passed
  • shared: 36 files, 363 passed
  • lint, format, typecheck, i18n parity, agent-team skills, and license headers passed
• pnpm run build — passed
• pnpm --filter @huabu/server exec vitest run src/modules/storage — 18 files, 335 passed against Disk, SQLite, and the shared port contracts
• pnpm --filter @huabu/server typecheck — passed