Upgrade to stable Effect v4 and fix reliability failures

[zackbart]

Pin Effect to stable 4.0.0 and fix 19 reproduced failures in request cancellation, OAuth, paused execution, storage, artifacts and the operator UI. Public deployment APIs stay Promise-based, and the eight MCP tools retain their contracts.

Doctor now refuses redirects before deployment credentials can reach another origin. Duplicate API tool names fail at construction. Unreadable successful Notion responses fail without retrying a dispatched write. Accepted OAuth rotations survive owner cancellation and share one persistence operation, while disconnect generations remain fenced.

The PR also records the Effect MCP, Schema and HTTP evaluation. The MCP compatibility gate is tracked in #622.

Fixes

• Worker catalog mutation ordering and request liveness.
• Token reservation rollback, callback transport cleanup and refresh-token persistence.
• Run-wide write budgets and the 4 MiB journal bound, with truthful known/unknown outcomes.
• File-storage rollback, arbitrary keys, exact TTL expiry and invalid snapshot quarantine.
• Artifact pagination, cancelled render checks, viewer themes, filter ordering, token-list mutation ordering and confined-shell reauthentication.

Validation

• npm run release:check passed on the integrated candidate: 4,890 tests passed, 56 skipped, and 52 Chromium browser tests passed.
• Typecheck, lint, unused checks, declarations, bundle budgets, examples and Docker/package smoke passed.
• Production dependency audit reported zero vulnerabilities.
• Regressions exercise Node and Workers, storage faults, concurrent resumptions, OAuth rotation races and out-of-order browser responses.
• Public providers:check remains unavailable because Vercel's categorized reference no longer matches the inventory reader. Tracked separately in Restore the complete Vercel public MCP inventory check #642.

Fresh Sol 6.1 review found no material issues on head dcd1e075. Independent Sol verification installed dependencies from the lockfile, passed 4,890 Node/Workers tests and 20 additional synthetic probes, reproduced the local Effect evaluation, and passed the build, declaration, bundle, examples, security and Docker/package stages. Its local Chromium launch was blocked by the macOS sandbox before any browser case ran. GitHub CI run 36878740424 passed the full release:check on that exact head, including browser tests.

Fixes #623. Fixes #624. Fixes #625. Fixes #626. Fixes #627.
Fixes #628. Fixes #629. Fixes #630. Fixes #631. Fixes #632.
Fixes #633. Fixes #634. Fixes #635. Fixes #636. Fixes #637.
Fixes #638. Fixes #639. Fixes #640. Fixes #641.