fix(drift): classify full live /models family wave + durable preview-exclude rule

[jpr5]

Problem

The daily model-family drift canary (src/__tests__/drift/models.drift.ts)
flagged the entire 2026-07-16 live GET /models wave as UNCLASSIFIED —
families present in neither includeFamilies nor excludeFamilies, which turns
the drift job red:

• OpenAI: 48 unclassified families (gpt-5.x tiers, Codex line, o-series,
  retired base models, image/video/moderation/search/deep-research surfaces)
• Anthropic: 10 (all Claude 4.x/5 point releases + claude-fable-5)
• Gemini: 45 (stable flash/pro tiers, plus preview/image/audio/tts/video/
  music/robotics/embeddings/Gemma surfaces)

Root cause of the recurrence: excludeFamilies is a hand-maintained enumerated
list with no rule for the churn-heavy families (previews, Gemma variants),
so each new one re-fires the canary as brand-new drift and needs a manual
registry edit — forever.

Fix

Classification (per resolved decisions):

• INCLUDE (text chat / reasoning aimock mocks): OpenAI gpt-5.x tiers + chat
  surfaces, the Codex line (gpt-5-codex … gpt-5.3-codex),
  gpt-5.6-luna/-sol/-terra, o-series; Anthropic all 4.x/5 point releases
  including claude-fable-5 (included ahead of a recorded fixture — intended,
  mirrors the forward-looking rationale of the exclude-by-rule patterns); stable
  Gemini flash/pro tiers.
• EXCLUDE: retired base (babbage/davinci/gpt-3.5-), embeddings, image
  (gpt-image/Imagen/nano-banana/gemini-*-image), audio/TTS, moderation, video
  (Sora/Veo), music (Lyria), robotics, deep-research, search, computer-use,
  antigravity, aqa.
• Moving aliases (chat-latest, gemini *-latest): excluded as non-family
  aliases — no existing include convention for them.

Durable hardening — two exclude-by-rule predicates:

• PREVIEW_FAMILY (/-preview(-\d+)?$/) — any family whose normalized key
  ends in -preview (optionally -preview-NN) is excluded by RULE.
  Previews are blanket-excluded until GA per policy: this intentionally
  silences the canary on new text previews (e.g. o1-preview, gpt-4.5-preview)
  until they drop the -preview suffix and go GA. There is no include-side
  escape hatch — aimock does not mock preview surfaces.
• GEMMA_FAMILY (/^gemma(-|$)/) — the open-weight Gemma line is out of
  scope (aimock does not mock it; it only rides the shared Gemini /models
  list). Replaces the two enumerated literals (gemma-4-26b-a4b-it,
  gemma-4-31b-it) so future Gemma variants auto-exclude with zero registry
  churn. Reversible by explicit handling if a Gemma fixture is ever added.

Removed the dead include-wins escape hatch: the prior isClassifiedFamily
had an "include wins" branch so an included family ending in -preview would
still be treated as included. No consumer distinguished include from exclude for
preview families, so the branch was dead code, and the two tests guarding it
were false-green (they used gemini-x-chat-preview, which matches
PREVIEW_FAMILY, so they passed even with the branch deleted). Both the dead
branch and the false-green tests are removed. Verified
include ∩ (preview ∪ gemma) = ∅, so removal is a functional no-op.

A family is now classified iff it is in include ∪ exclude, or matches an
exclude-by-rule pattern (preview, gemma). Implemented at the classification
layer, not the normalizer, so preview/gemma family keys stay distinct and the
canary keeps surfacing genuinely-new non-preview, non-gemma families. The
builder/fixture cross-check uses the same isClassifiedFamily predicate so the
two classification surfaces cannot drift apart.

No version bump — package.json untouched (versioned separately).

Red → Green proof

include ∩ (preview ∪ gemma) = ∅ — verified empirically before removing the
include-wins branch: no family in any includeFamilies set matches
PREVIEW_FAMILY or GEMMA_FAMILY, so the removal cannot mis-classify any
included family.

GEMMA_FAMILY rule (new genuine red-green test):

RED — against the pre-fix registry (literal names only, no GEMMA_FAMILY):

 × isClassifiedFamily / PREVIEW_FAMILY > classifies a future Gemma variant by rule
   → Cannot read properties of undefined (reading 'test')   # GEMMA_FAMILY did not exist

A synthetic gemma-9-foo (normalizes to itself — no numeric-only build tag to
strip) was on neither include nor exclude and was unclassified under the old
literal-names-only set.

GREEN — with the pattern:

 ✓ isClassifiedFamily / PREVIEW_FAMILY > classifies a future Gemma variant by rule
 ✓ gemma families are excluded by rule > a future Gemma variant auto-excludes with zero registry edits

Full drift + unit suite (after the fix):

 Test Files  16 passed | 8 skipped (24)     # pnpm test:drift — 0/0/0 unclassified
      Tests  98 passed | 55 skipped (153)
 Test Files  158 passed (158)               # pnpm test — full unit
      Tests  4670 passed (4670)

The full-wave regression suites (OpenAI/Anthropic/Gemini) still assert zero
unclassified against the real unclassifiedFamilies pipeline. The two
false-green INCLUDE-WINS tests are gone; no tautological test replaces them.

format:check / lint / tsdown build / commitlint (header ≤100) all pass on
the changed files.

[pkg-pr-new]

Open in StackBlitz

npm i https://pkg.pr.new/@copilotkit/aimock@313

commit: 043410b