test: declare the identity sweep's skips instead of letting them shrink it

[sroussey]

Stacked fix onto #741 (claude/coverage-dist-bundle-fix-ew0vj8), which is the base of this PR — not main.

Stacked on top of #817, which it depends on: the two-sided base-class assertion below reads the canonical WORKGLOW_TEST_TARGET that #817's shared.env hands to the workers. Merge #817 first; until it lands, this PR's diff also shows #817's commit. The changes authored here touch exactly one file (packages/test/src/test/util/PublishedEntryIdentity.test.ts) — zero file overlap with #817.

MEDIUM — the sweep skips the exact regression it exists to catch

In PublishedEntryIdentity.test.ts, a candidate whose ai-runtime module exposes no register*Inline export is pushed onto withoutInlineRegistrar and continued, contributing no assertion at all. The only place that surfaced was a message that prints on failure — i.e. never, on a green run.

Counted on this tree: 16 workspace packages publish both ./ai and ./ai-runtime (anthropic, cactus, chrome-ai, deepseek, google-gemini, huggingface-inference, huggingface-transformers, llamacpp-server, mlx, node-llama-cpp, ollama, openai, openrouter, stable-diffusion-server, tf-mediapipe, xai). Every anti-vacuity bound was > 4. So 11 of 16 providers could drop their registrar and leave the file green.

Measured, not reasoned. Renaming registerAnthropicInline in providers/anthropic/src/ai/registerAnthropicInline.ts and rebuilding that provider:

# built bundle after the rename
matching exports: []

$ WORKGLOW_TEST_TARGET=dist npx vitest run … PublishedEntryIdentity.test.ts
Test Files  1 passed (1)
     Tests  6 passed (6)          # GREEN — @workglow/anthropic silently gone from the sweep

With this PR's assertions and the rename still applied, the same run is RED, naming the package in both new checks:

AssertionError: expected [ '@workglow/anthropic', '@workglow/mlx' ] to deeply equal [ '@workglow/mlx' ]
  a runtime entry exporting no register*Inline is checked by nothing — declare it in
  NO_INLINE_REGISTRAR with the reason, or restore its registrar

AssertionError: expected [ '@workglow/cactus', …(13) ] to deeply equal [ '@workglow/anthropic', …(14) ]
- "@workglow/anthropic"

Test Files  1 failed (1)
     Tests  2 failed | 5 passed (7)

The rename was reverted and the provider rebuilt (matching exports: [ 'registerAnthropicInline' ]).

Changes

• NO_INLINE_REGISTRAR, declared beside NEEDS_NATIVE_RUNTIME. Verified sole member today: @workglow/mlx, which exports registerMlx — no Inline suffix — because MlxProvider.isAvailable reports false until an mlx-lm runtime is bundled, so there is nothing to register inline. withoutInlineRegistrar.sort() is asserted equal to its keys, so a provider that drops its registrar in a refactor fails instead of silently shrinking the sweep.
• The exemption-pinning test now covers both maps — either one removes a candidate from the checks — asserting every key still names a real candidate and every reason is longer than a word, with the map name in the assertion label.
• MINIMUM_RUNTIME_CANDIDATES (12) replaces candidates.length > 4: deliberately below today's 16, since a provider may legitimately be removed, but far enough above zero that a typo in the walk fails rather than yielding a short list. checkable.length becomes an equality — candidates.length - Object.keys(NEEDS_NATIVE_RUNTIME).length — rather than a second floor, so a candidate falling out of the sweep for any undeclared reason fails.
• The providers.size > 4 block becomes two statements a shrunken sweep cannot satisfy: the registered package set equals the checkable set minus the declared no-registrar packages, and each registration is checked for providerNames.length >= 1, collected into an offenders array so one no-opping registrar reports itself instead of being averaged away by the fourteen that worked.

Two-sided base-class assertion

expect(new Set(PUBLISHED_BASE_CLASSES).size).toBeLessThanOrEqual(2) is satisfied by 1 or 2, so it could not tell a real dist run from a source run mislabelled as one — nor from a stubbed dist re-exporting src. It is now target-keyed:

• dist ⇒ exactly 2 — @workglow/ai builds . and ./worker as separate bun build invocations, so two distinct bundles really loaded;
• source ⇒ exactly 1 — both specifiers resolve through one underlying module, so the resolution plugin really attached.

This is the in-process proof that the modules under test are bundles rather than src. The variable it reads is the one resolveTestTarget already validated in vitest.config.ts (#817), so === "dist" cannot silently mean "source" for a typo the way an unvalidated read would.

Verification (actual output)

$ npx vitest run --config vitest.config.ts …/PublishedEntryIdentity.test.ts
Test Files  1 passed (1)          # source target: asserts exactly 1 base class
     Tests  7 passed (7)

$ WORKGLOW_TEST_TARGET=dist npx vitest run --config vitest.config.ts …/PublishedEntryIdentity.test.ts
Test Files  1 passed (1)          # dist target: asserts exactly 2
     Tests  7 passed (7)

$ bun run test:vitest:dist
Test Files  537 passed | 3 skipped (540)
     Tests  6273 passed | 71 skipped (6344)

$ bun run typecheck:budget       # typecheck-budget: OK (38 packages within budget)

---

Generated by Claude Code

[github-actions]

Coverage Report

Status	Category	Percentage	Covered / Total
🔵	Lines	71.82%	29671 / 41311
🔵	Statements	70.63%	31711 / 44891
🔵	Functions	72.33%	6136 / 8483
🔵	Branches	61.19%	16011 / 26162

File Coverage

No changed files found.

Generated in workflow #3168 for commit 89e8c86 by the Vitest Coverage Report Action