fix: preserve usage for dynamically loaded AI SDK V2 providers - #50414

Issue for this PR

Closes #50338

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Provider.getLanguage() is typed as returning LanguageModelV3, but a dynamically loaded provider can hand back a runtime LanguageModelV2 instead. session/llm.ts always passed that model through wrapLanguageModel(). Looking at wrapLanguageModel's implementation in the ai package, the object it returns is hardcoded to specificationVersion: "v3" and calls the wrapped model's doGenerate/doStream directly — it doesn't itself do any V2→V3 conversion. AI SDK has its own compatibility shim for this (asLanguageModelV3, used inside resolveLanguageModel), but that shim only runs when streamText sees a model whose specificationVersion is still "v2". Since wrapLanguageModel already relabels it "v3" before streamText ever looks at it, the shim never fires and the V2 usage numbers (a different shape than V3) never get converted — they come back undefined.

That matters beyond the usage indicator: session/overflow.ts uses that usage to decide when to auto-compact, so with it undefined compaction never triggers and long sessions on these providers just keep growing until the provider hard-rejects the request.

Fix: check the runtime specificationVersion before wrapping. V3 models go through wrapLanguageModel() exactly as before. V2 models skip the wrapper — the message transform that used to run inside transformParams gets applied directly to the ModelMessage[] input instead, and the raw V2 model is passed straight to streamText, so AI SDK's own compat layer does the conversion the way it's meant to.

How did you verify your code works?

• Wrote a standalone script against wrapLanguageModel/streamText directly (not opencode code) with a hand-built V2 model to confirm the mechanism: wrapping it drops usage, passing it unwrapped preserves it. This matches the reporter's own differential probe in the issue.
• Added a regression test in test/session/llm.test.ts with a small fixture provider (test/session/fixtures/fake-v2-provider.ts) that implements LanguageModelV2 and is loaded via npm: "file://...", the same mechanism any dynamically loaded provider uses. It goes through the real Provider.getLanguage() → LLM.Service.stream() path. Checked it fails on the pre-fix code (usage undefined) and passes with the fix.
• bun run typecheck in packages/opencode — clean.
• bun test test/session/llm.test.ts — 31/31 pass.
• bun test test/session/ (full directory) — 413 pass, 0 fail, no regressions.

Screenshots / recordings

Not a UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR