deps(ai): migrate the Vercel AI SDK from 6 to 7

[mrsibe]

What

Upgrade the AI SDK from 6 to 7, on top of the already-merged 5 → 6 PR (#184).

Package	Before	After
ai	6.0.297	7.0.123
@ai-sdk/provider	3.0.18	4.0.20
@ai-sdk/openai	3.0.121	4.0.81
@ai-sdk/anthropic	3.0.125	4.0.69
@ai-sdk/google	3.0.129	4.0.86
@ai-sdk/openai-compatible	2.0.80	3.0.60

The codemod, and why it looked "stuck"

npx @ai-sdk/codemod v7 ignores its positional path argument and scans the repository root. Its ignore list covers node_modules, dist and build, but not out/ — the bundled renderer output (6 MB of index-*.js plus the 1.3 MB pdf worker). jscodeshift parsed that once per codemod, ~20 times, which reads as a hang.

Removing out/ and dist/ first takes the whole run from "stuck" to one minute. Worth knowing for the next major.

After that, the codemod touched exactly one line: createGoogleGenerativeAI → createGoogle. KnowNote does not use the other renamed surfaces (fullStream, onFinish, telemetry, tools, custom providers).

⚠️ Silent behavioural regression: v7 drops system messages and returns an empty answer

This is not a compatibility nit. It is the kind of change that makes RAG look like it
broke for no reason: no exception, no error event, no non-200 — just a blank answer.
It is recorded here and in commit 87c33fc deliberately, so it is findable later.

AI SDK 7 rejects { role: 'system' } messages inside messages by default — and it does not throw. Verified directly:

default:               OK -> ""          # empty answer, no error
allowSystemInMessages: OK -> "hi"

KnowNote puts the retrieved context (RAG) and the title-generation prompt into the history as a system message (chatHandlers.ts ×2, noteHandlers.ts). Left alone, every RAG turn would silently produce a blank answer — worse than a crash, because nothing surfaces.

ModelClient.streamChat now passes allowSystemInMessages: true. Every message it sends is built by this app, not by the user, so the untrusted-injection case that the default guards against does not apply. A regression test asserts the system message actually reaches the provider's prompt.

Also done: stateless toUIMessageStream

result.toUIMessageStream(...) is deprecated in v7 — it works but warns every turn and is slated for removal in v8. Moved to the stateless helper, same options:

toUIMessageStream({ stream: result.stream, sendReasoning, onError, messageMetadata })

The messageMetadata usage-on-finish wiring is unchanged, so the forwarded event stream is identical — which is exactly what the chat-stream and UI-message tests cover.

Verified (Linux x64, WSL2)

Command	Result
npm run typecheck	clean
npm test	471/471 pass (470 + the system-message regression test)
npm run build:unpack && npm run smoke:packaged	29 checks pass

The packaged smoke matters here because v7 is ESM-only: ai and the providers are devDependencies that rollup bundles into out/main. The build succeeds and the real packaged app starts and self-checks.

Also checked rather than assumed:

• streamObject still exists in v7 (deprecated, not removed), so quiz / Anki / mind-map keep working.
• LanguageModelUsage still exposes the numeric inputTokens / outputTokens / totalTokens this app reads.
• readUIMessageStream and MockLanguageModelV3 are still present.

Known remaining deprecations (not fixed here)

• streamObject is deprecated in favour of streamText + Output.object. Rewriting QuizService, AnkiCardService and MindMapService is a behaviour-sensitive change (partial-object streams, progress reporting) and belongs in its own PR, not a version bump.
• The 'openai-compatible' providerOptions key is still read but should become 'openaiCompatible'. It only logs a warning; renaming changes which key the provider reports providerMetadata under, so I left it for a focused follow-up.

Rollback

Revert the branch. No schema, storage format, or persisted data changes.