test(drift): contain the DATA_FROZEN collection blast radius, and make the mirror guard catch the deprecation .sort()

[jpr5]

Two post-merge review findings on the drift guards, fixed forward off fresh main. Both arrived after PR #343 had already merged (d835753), so this is a follow-up rather than an amendment to anything. Tests only — no product code, no version-bearing file, no scripts/, no .github/.

M1 — DATA_FROZEN.members was evaluated at module scope

FROZEN.source was made a thunk for one reason: so that one renamed or reshaped surface fails its OWN case instead of throwing during collection and taking the whole file with it. DATA_FROZEN.members, twelve lines below, kept the module-scope evaluation — the identical defect, in the twin of the surface we fixed. Vitest resolves a renamed export to undefined rather than failing the import, so [...undefined] at module scope throws at collection time.

RED (pre-fix) — rename gaRealtimeModels in voice-models.ts, run logic-pin.test.ts:

TypeError: gaRealtimeModels is not iterable
 ❯ src/__tests__/drift/logic-pin.test.ts:433:18
    members: [...gaRealtimeModels].sort(),

 Test Files  1 failed (1)
      Tests  no tests

One rename zeroed the file: 12 source pins, 7 behavioural anchors, and all 9 data pins. Baseline for that file is 29 passing.

GREEN (post-fix) — same rename, blast radius contained:

 FAIL  ... > freezes detectVoiceModelDrift
 FAIL  ... > freezes gaRealtimeModels membership
      Tests  2 failed | 30 passed (32)

Both failures are surfaces the rename genuinely moved (detectVoiceModelDrift's body references the renamed const). The rename was then edited back out: 32/32 pass.

Checked for a third module-scope surface in that file: there are exactly three module-scope bindings — srcCache (an empty Map), FROZEN (all thunks), DATA_FROZEN (now all thunks). No third defect.

While in the file, pinned FORWARD_LOOKING_FAMILIES per provider (including the two empty sets). A family listed there is dropped from the sync mirror's deprecation candidates entirely — no removal proposal, no needs-human note — so adding an id was a one-line way to stop reporting a genuinely retired family, and no guard watched it. Mutation-checked: seeding the openai set reddens that pin and nothing else.

M2 — the mirror-equivalence guard did not catch the deprecation .sort()

There are two .sort()s in the drift-sync mirror, one per mirrored function, and only one was covered. Every deprecation fixture yielded at most ONE missing family (one candidate is trivially in order), and the sole two-candidate case asserted with order-insensitive toContain.

RED (pre-fix) — drop detectDeprecatedFamiliesForSync's .sort():

src/__tests__/drift-sync-mirror-equivalence.test.ts   6 passed (6)
src/__tests__/drift-sync-core.test.ts                34 passed (34)

Fully green with the divergence in place.

This corrects the record on commit c89cf6a ("make the mirror-equivalence guard actually catch all three named divergences"), which landed on main in #343. That claim was true of two of the three, not three: the fail-closed floor and the NON_MODEL_TOKENS check were genuinely covered, and the .sort() fixture it added covers unclassifiedFamiliesForSync's sort — a different .sort() in a different function. The deprecation copy's sort was uncovered. A commit message on main cannot be rewritten, so the correction is recorded here and in the file's docstring, which now states there are two sorts and that each needs its own order-distinguishing fixture.

GREEN (post-fix) — the new fixture removes two families whose registry insertion order is not alphabetical, so sorted and unsorted missing differ, and asserts the whole result with order-sensitive toEqual:

AssertionError: expected { status: 'checked', …(1) } to deeply equal { status: 'checked', …(1) }
- Expected
+ Received
-       "family": "gpt-4.1",
+       "family": "gpt-4o",
...
-       "family": "gpt-4o",
+       "family": "gpt-4.1",

Restored: 8/8 pass. It also pins the order absolutely, so dropping the sort from both copies at once (which keeps them equivalent) still reddens — verified. The pair is derived from the registry, not hard-coded, and firstUnsortedPair throws if the registry ever becomes fully sorted rather than letting the fixture go quietly vacuous.

Re-verified the other two named divergences really are caught, post-fix:

mutation to the mirror	result
drop NON_MODEL_TOKENS.has(...) line	2 failed | 6 passed
weaken the fail-closed floor to if (false)	1 failed | 7 passed
drop unclassifiedFamiliesForSync's .sort()	1 failed | 5 passed
drop detectDeprecatedFamiliesForSync's .sort()	1 failed | 7 passed (was 6 passed)

Smaller items taken in the same pass

• mirror: the || NON_MODEL_TOKENS.has(id) half is unexercised. The comment claimed the fixture protects the token check; it protects the line as a whole. Reducing the mirror's line to has(family) alone keeps the file green, because the sole token normalizes to itself. Narrowed the comment to what is true rather than widening the code.
• mirror: the "real isFamilyStillReferenced" title was false — that fixture's missing set is empty, so it never reaches the reference scan. Title corrected, and the case now pins { status: "checked", candidates: [] } so the equivalence cannot be satisfied by both copies going equally wrong. The new two-missing-families fixture is what exercises the real scan (it passes no isReferenced).
• mirror: gemini was never exercised — the two exclude-by-rule lanes (PREVIEW_FAMILY, GEMMA_FAMILY) had no coverage on either copy. Added.
• mirror: fixture identity on failure — the seven-fixture loop now names the listing in its assertion message; also dropped the re-declared local Provider in favour of the mirror's own exported type.
• text-drift.test.ts: a toEqual([]) that a neutered detector also satisfies. It fed in the bare registry keys (which are excludeFamilies.openai by definition, so that half could never contribute) and passed under unclassifiedFamilies → return []. Now dated-only, plus a negative control of the same shape that must report. Verified: it reddens under that mutation, where before it stayed green.
• text-drift.test.ts: the generated id did not exist. The loop derives ids from the registry keys, which are seeded through normalizeModelFamily, so text-embedding-ada-002 is stored as text-embedding-ada and the generated id was text-embedding-ada-2026-07-28. Added the real bare and dated forms, which need the -002 build tag and the date to normalize in composition.

Nothing deferred.

Gates

• pnpm build — clean
• pnpm test — 4764 passed / 4765, 0 skipped. The one failure is the known timing-replay.test.ts wall-clock flake (replaySpeed 2.0), which passes 5/5 isolated and also failed on unmodified origin/main in this worktree. Six new tests (3 + 2 + 1), consistent with the 4759 baseline.
• pnpm lint — clean · prettier --check . — clean
• Nothing re-pinned: git diff origin/main..HEAD | grep -E '^[-+].*pin:' shows only the three NEW FORWARD_LOOKING_FAMILIES pins and the DATA_FROZEN type signature. No existing pin line was removed or changed.
• pnpm test:drift includes only *.drift.ts — untouched, so no live canary is in play and no provider key was used.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@351

commit: 4f845ae