test(exports): widen both export guards one directory down, and fix what they find

[sroussey]

Stacked on #717 (base: claude/wonderful-turing-rjtcnx-ai-types). The manifest work in #717 is not touched here — this is the same defect class surviving one directory below where #717's new guards look.

Both guards landed in #717 examine only the layer they were written for:

• ExportBarrelParity keyed on the hard-coded src/ai/index{,.browser}.ts pair, so it never saw src/ai/runtime.browser.ts — the barrel a browser consumer gets under @workglow/<vendor>/ai-runtime.
• ExportTypesPairing's duplicateBrowserEntryViolations collected pairs with a flat readdirSync(srcDir), so it saw only the entry layer the manifests name.

Two commits, deliberately not squashed

The A-then-B ordering is the evidence the widening bites. Commit A widens the guards only and is red, reporting exactly two defects and nothing else:

providers/ollama/src/ai/runtime.ts exports
  * from "./common/Ollama_StructuredGeneration", * from "./common/Ollama_TextGeneration"
  but providers/ollama/src/ai/runtime.browser.ts does not

providers/openrouter/src/ai/runtime.browser.ts is identical to
  providers/openrouter/src/ai/runtime.ts and names only relative specifiers
  ("./common/OpenRouter_Client", "./common/OpenRouter_EffortPolicy",
   "./registerOpenRouterInline", "./registerOpenRouterWorker"),
  so both entries resolve the same module graph and the declaration split is
  nominal — re-export the node entry from it instead of duplicating it

Commit B fixes those two sources and turns it green.

What the widening needed

The recursive scan takes ExportBarrelParity from 5 pairs to ~53, and three things had to come with it:

• Specifier keys strip a trailing .browser. Without it every runtime pair reports as total drift (* from "./common/Ollama_Client" against * from "./common/Ollama_Client.browser") and the guard says nothing. The residual is documented in the source: this makes a star-export comparison across a .browser sibling nominal — and the recursive scan is exactly what closes it, since Ollama_Client.browser.ts vs Ollama_Client.ts is itself a compared pair now.
• A pure sibling re-export is skipped. A browser file whose entire surface is * from "./<nodeStem>" IS the node surface, and it is the shape fix(providers): point each export condition's types at its own declaration #717 wants — the widened scan now reaches llamacpp-server and stable-diffusion-server, which would otherwise report the node file's every other statement as drift.
• INTENTIONAL_NODE_ONLY is rekeyed from package dir to browser FILE path, since a package now contributes several pairs. Both consumers (drift loop, staleness loop) key on pair.browserPath.

The vacuous-pass assertion is removed rather than adjusted. expect(pairs.map(…).sort()).toEqual([...INTENTIONAL_NODE_ONLY.keys()]) is flatly wrong at 53 pairs against 5 pins, and any form of it still means a correct new provider fails until somebody registers it as needing no exemption. Its two real roles are stated separately instead: pairs.length > 40 (the scan found the tree) and pairs.some(p => p.browserPath.includes("/src/ai/")) (and it recursed into it). The staleness test is kept, retargeted to the path.

Scope stays at providers/ deliberately, and the doc comment says why: extending to packages/ pulls in packages/tasks/src/task/image/imageTextRender.browser.ts, a genuine implementation split whose browser file exports one factory against the node module's fifteen names — a fifteen-name exemption buying nothing.

For ExportTypesPairing, the relative/bare distinction is unchanged and is still the whole rule. Verified under the recursive scan: all nine packages/workglow bare-specifier shims still classify BARE and stay unreported. The > 9 floor is kept (they are still the negative case) and a some(p => p.browserPath.includes("/src/ai/")) assertion is added so a regression to the flat scan fails loudly.

The source fixes

ollama is a source-compatibility fix, not a cosmetic one. Before #717 the browser types condition pointed at the NODE declarations, so a browser consumer importing createOllamaTextGenerationStream from @workglow/ollama/ai-runtime compiled fine and got undefined at runtime. With #717's accurate types it becomes a hard TS2305 for anyone with customConditions: ["browser"].

Neither module has a .browser variant and neither imports a node builtin; both are already compiled into the browser bundle via Ollama_JobRunFns.browser. Verified against a rebuild: ai-runtime.browser.js gains no code — the only content delta is the two names in the entry's export list, plus a module reordering because they are now direct entry re-exports.

Every other provider's runtime.browser.ts was audited against its runtime.ts: deepseek, openai, openrouter and xai are in parity modulo the .browser specifier suffix. ollama was the only omission. (The ai-runtime.browser.ts entry layer was audited too; llamacpp-server and stable-diffusion-server differ only in being the intended sibling re-export.)

openrouter was a byte copy naming only relative specifiers, so both halves already resolved the same module graph and the split was nominal. It now re-exports its node peer, reusing the wording #717 established in providers/llamacpp-server/src/ai-runtime.browser.ts. Rebuilt ai-runtime.browser.js is byte-identical (modulo debugId).

Verification

check	result
bun scripts/test.ts util vitest after A	2 failed | 814 passed — exactly the two above
bun scripts/test.ts util vitest after B	55 files, 816 passed, 10 skipped
ExportBarrelParity + ExportTypesPairing directly	62 passed
BunExportConditions.test.ts	3 passed — the pinned "bun" condition set is untouched (no package.json is modified by either commit)
bun run build:types	41 successful
bunx turbo run build-js build-types --filter=@workglow/ollama --filter=@workglow/openrouter	10 successful
bun run format	clean

Notes for whoever merges

Rebase resolution rule. This branch was rebased onto origin/main before the fixes. main had added export { <vendor>EffortPolicy } from "./common/<Vendor>_EffortPolicy"; to both barrels of openai / xai / deepseek / openrouter while #717's own commit added export * lines to index.browser.ts in the same region. deepseek and xai auto-merged; openai and openrouter conflicted and were resolved as the UNION of both sides in BOTH files. Taking one side's EffortPolicy line in index.ts but not index.browser.ts — or dropping a branch-restored export * — silently recreates exactly the asymmetry ExportBarrelParity exists to catch. The resolution was verified with the guard itself (a dropped line surfaces as a new <vendor>EffortPolicy from "./common/<Vendor>_EffortPolicy" entry in the drift list); no such entry appeared.

Diff size. Because this branch is rebased onto main and the base branch is not, GitHub computes the diff from an old merge base and shows ~117 commits. The change actually under review is the two commits at the tip; git diff <PR-717-tip-as-rebased>..HEAD is 4 files. Rebasing #717 onto main collapses this to those two commits.

---

Generated by Claude Code