feat(runtime): upgrade Vercel AI SDK to v7

[YayoiNanoka]

Summary

Upgrade Maka from Vercel AI SDK 6 to AI SDK 7 while preserving the existing multi-provider model integration, Tool Calling Agent Loop, permission handling, context management, retries, and usage accounting.

This PR also removes obsolete archived sandbox development plans and updates the archive index.

Changes

• Upgrade ai from 6.0.185 to 7.0.31.
• Upgrade the provider packages:
  • @ai-sdk/openai to 4.0.16
  • @ai-sdk/anthropic to 4.0.16
  • @ai-sdk/google to 4.0.18
  • @ai-sdk/cohere to 4.0.11
  • @ai-sdk/openai-compatible to 3.0.12
• Migrate provider contracts from LanguageModelV3 to LanguageModelV4.
• Replace createGoogleGenerativeAI with createGoogle.
• Migrate AI SDK APIs:
  • system → instructions
  • stepCountIs() → isStepCount()
  • fullStream → stream
  • experimental_repairToolCall → repairToolCall
  • experimental_context → runtimeContext / toolsContext
• Adapt Maka's prepareStep composition to AI SDK 7's carry-forward semantics.
• Prevent steering messages, compact results, and tool-result pruning from being applied more than once across steps.
• Update usage accounting for AI SDK 7's cumulative usage semantics.
• Preserve per-step input-token tracking for context-window reporting.
• Migrate user images, MCP images, and Computer Use screenshots from image-data to V7 file content parts.
• Migrate standalone generateText() consumers in runtime, headless, and desktop.
• Upgrade test infrastructure to MockLanguageModelV4 and V4 stream/usage contracts.
• Add an upgrade checklist documenting completed work and remaining live-provider coverage.

Agent Loop compatibility

Maka continues to use AI SDK's streamText() Tool Calling loop while retaining Maka-owned tool registration, permission checks, dynamic tool activation, context compaction, retries, persistence, and telemetry.

The migration specifically verifies:

• explicit and unbounded step stopping;
• deferred tool activation;
• prepareStep message carry-forward;
• permission allow/deny/wait/abort;
• overflow and transport retry behavior;
• prevention of side-effect replay;
• cumulative usage and cost accounting.

Validation

The following completed successfully:

• npm ci
• npm run clean
• npm run build
• npm run typecheck
• npm run lint
• npm test
• Runtime full test suite
• All workspace test suites
• Deprecated V3/V6 API scan
• git diff --check
• Pinned cua-driver v0.7.1-maka.2 artifact integrity check

Live provider smoke tests

DeepSeek official API through the OpenAI-compatible adapter:

• deepseek-chat
  • text streaming passed;
  • two-step Tool Calling passed.
• deepseek-v4-pro
  • text streaming passed;
  • reasoning streaming and reasoning usage passed;
  • two-step Tool Calling passed;
  • cumulative usage and cached-input usage were observed.

Known validation gaps

• The configured tokenhub/OpenAI gpt-5.5 request reached /v1/responses, but the provider returned HTTP 402 due to insufficient account quota.
• Anthropic, Google, and Cohere live smoke tests were not run because no credentials were available.
• The isolated Electron Computer Use observe-only E2E was executed, but deepseek-v4-pro is classified as a non-vision model. Maka therefore intentionally removes the screenshot-returning maka_computer tool before the request, so this model cannot satisfy the Computer Use qualification gate.
• npm ci reports 4 audit findings (2 low and 2 high). No forced audit upgrade is included in this PR.

Documentation cleanup

Remove obsolete archived sandbox implementation plans that no longer describe the current architecture, and update docs/archive/README.md accordingly.