[callumalpass]

Summary

Migrate Writer to its existing SDK-only backend so weekly production-verified repins need no backend migration. UI and i18n are now exactly 0.1.0-beta.148.

Legacy callers and their replacements

• No-environment/default app entry: removed LegacyApp, ConnectBackend, its Connect application session, gates/picker and unused sign-in/popup preparation. All ordinary builds use the SDK application entry and NativeWriterBackend; absent environment defaults to production, as the manifest builder does. Explicit environment/origin checks still refuse mismatches. No fallback or opt-in backend flag.
• Demo, large demo and performance fixture: moved from connect-testing / synthetic legacy queries to the retained SDK's MemoryReplica, using the same native backend. Fixture-only catalog and array-type derivation fill gaps in MemoryReplica, not in the production SDK. Browser fault injection now uses public SDK/replica APIs.
• Workspace/outline tests: use actual SDK leases and record sessions. Removed tests specific to the deleted legacy query transport; native tests and browser coverage remain.
• Shared JSON views: removed legacy type imports. Writer's JSON draft/session port stays because DraftStore consumes persisted browser backups. Connected namespaces remain collection IDs; the demo namespace remains demo. This is application data, not a protocol compatibility shim.

The SDK original sdk-eaa4e697-production-beta141.tgz is unchanged (SHA256 e64771a90bed2fb3837d4b65df8930fcd124d3059fcb0bcbaf9d64dbfb148205). No SDK changes or repacking. The matching protocol original remains for manifest validation and publisher custody checks. Removed the beta.125 Connect dependency/override, connect-dev/testing dependencies, both query-patched archives and their patch/pin sidecars. No unrelated dependencies changed.

PR #16 was inspected. Its structural workspace session interface is already present on main and retained; its older SDK, proposed control plane, custom session engine and opt-in dual-backend design are not revived.

Flow corrections found during migration

• Collection resets invalidate expanded annotation bodies as well as discovery metadata.
• Recovery problems remain visible, and Retry save delegates to SDK flush/receipt recovery without resubmitting an unknown operation.
• “Use as manuscript” preserves already-typed notes. The pinned SDK's contract promotion only accepts untyped notes, so this named persisted-record flow uses one ordinary SDK CAS update and verifies the contract readback. Existing mapped titles, native sibling values, inferred/explicit types, body and record ID are retained; malformed explicit type values refuse before submission. Regression tests cover these cases.
• Kept the existing error boundary/draft-download behavior in the shared shell; removed obsolete catalogue entries and regenerated all ten locales without translation services.

Publisher inputs

config/native-publish/production.json binds npm UI beta.148's exact version, SHA512 SRI and tarball SHA256. Native intake supports that registry pin and checks its frozen importer/resolution and installed package identity; SDK/protocol complete-member and producer provenance/trust checks remain intact. The npm UI install is SRI-verified, rather than claiming an archived UI member inventory. Tamper tests cover version, SRI and protected-hash substitutions. Synthetic CI-unit verification uses the current production inputs without hard-coded member counts that would obstruct later SDK repins.

Before a future authorized production publication, the protected writer-production environment's NATIVE_UI_SHA256 must match the new approved value 56f51e9dce7cd83149fe99efbbbdb43203288f724ed6b6e019886e1c3d65ad75. This PR does not mutate GitHub environment variables or release branches.

Historical staging package/lock snapshots intentionally remain unchanged: they are frozen publisher-test inputs, not Writer's active dependency graph. The separately qualified staging publisher still requires its own staging original intake; this PR does not retarget the production SDK to staging.

Verification completed before opening this PR

Finite pnpm checks/builds used the repo's pinned pnpm 11.15.1 through rpnpm (remote build runner); browser/local-service tests ran locally. Node 22.22.0.

• pnpm install --lockfile-only and pnpm install --frozen-lockfile — passed; only the requested dependency changes in the lock.
• rpnpm check — passed strict ten-locale completeness/composition/source checks, typecheck and all normal suites: 646 passed, 7 existing skips (core 65 passed/4 skipped; app Node 80/2; app Vitest 274/1; root Node 227/0). Existing skips are optional Pandoc parity/materialization, sibling Reader/contracts checkout comparisons, and ZIP integration coverage; the real browser export tests also passed.
• rpnpm --pull apps/writer/dist build — passed, including manifest generation/validation and optimized production build. rpnpm ... build:production also passed.
• rpnpm --class build test:publisher:ci-unit — 282 passed, zero skipped, including original/provenance/installer/forgery tests and the actual current production intake.
• rpnpm --pull apps/writer/dist --filter @mdbase-writer/app build:editing-test and local test:browser:editing — passed optimized React editing, cursor movement and undo/redo with a real Preview canvas.
• Cold Vite optimizer cache, local pnpm dev: test:browser — 56 baseline/reliability scenarios passed, including real PDF/DOCX/ZIP exports, typed-note adoption, annotations, conflicts, drafts, failures and recovery.
• Local test:browser:i18n — passed all ten locales, language negotiation/persistence, formatting, CodeMirror and failed-chunk retention.
• Local test:browser:a11y — passed 375px/1280px axe and keyboard/dialog focus scenarios.
• build:feedback-test plus local loopback preview and test:feedback — passed desktop/mobile feedback; all delivery intercepted.
• Opt-in migrated SDK benchmark (rpnpm --class build --filter @mdbase-writer/app test --config scripts/perf/vitest.config.ts) — passed.
• git diff --check and empty pnpm why for connect/connect-dev/connect-testing — passed.

Independent static review completed; its adoption preservation findings were fixed directly with regression tests. No remaining migration blocker identified. These are deterministic/demo checks, not live connected-collection or native-device acceptance.

No merge, release/publish workflow dispatch, publication or deployment.