fix(playground): pull the speculative tier - quantized framing degrades output

Measurement behind the tier (interscript-ml RESULTS.md 2026-09-12):
dynamic-int8 decoder graphs quantize activations per fed tensor, so
single-step and batched decode framings produce materially different
numerics - the batched-verifier output lost a word on a canonical
test row, and runtime acceptance measured 0.46 vs the probe's 0.99
(uniform-framing artifact). int4 CPU kernels are also slower than
int8 here, so the tier had no speed story either. SpeculativeModel
stays in the npm API as measurement infrastructure with the
constraint documented; the playground serves only the verified tiers.

Author: ronaldtse

e2e/playground-spec-live.spec.ts

@@ -27,9 +27,6 @@ import catalogue from "../data/maps-catalogue.json"
           <option value="">Off — deterministic maps only</option>
           <option value="ara-diac-layerdrop-1.0-int4">Arabic lite (int4, 95 MB, downloads once)</option>
           <option value="ara-diac-small-2.1-int8">Arabic 2.1 (int8, 264 MB, downloads once)</option>
-          <option value="spec:ara-diac-small-2.1-int8">
-            Arabic 2.1 + speculative (int4 drafts it, 359 MB, downloads once)
-          </option>
         </select>
 
         <label for="filter">System</label>
@@ -168,10 +165,6 @@ import catalogue from "../data/maps-catalogue.json"
     if (!voc) {
       return `import { transliterateAsync } from "interscript"\n\nconst out = await transliterateAsync(\n  "${code}",\n  "…",\n)`
     }
-    if (voc.startsWith("spec:")) {
-      const verifier = voc.slice(5)
-      return `import { transliterateAsync } from "interscript"\nimport { imf } from "interscript/ml"\n\nconst drafter = await imf.IMFModel.fromZipBytes(\n  (await imf.resolve("ara-diac-layerdrop-1.0-int4")).bytes,\n)\nconst verifierModel = await imf.IMFModel.fromZipBytes(\n  (await imf.resolve("${verifier}")).bytes,\n)\nconst spec = new imf.SpeculativeModel(drafter, verifierModel)\nconst vocalized = await spec.translate("…")\nconst out = await transliterateAsync(\n  "${code}",\n  vocalized,\n)`
-    }
     return `import { transliterateAsync } from "interscript"\nimport { imf } from "interscript/ml"\n\nconst resolved = await imf.resolve("${voc}")\nconst model = await imf.IMFModel.fromZipBytes(resolved.bytes)\nconst vocalized = await model.translate("…")\nconst out = await transliterateAsync(\n  "${code}",\n  vocalized,\n)`
   }
   $("snippet").textContent = snippetText()
@@ -189,10 +182,7 @@ import catalogue from "../data/maps-catalogue.json"
   async function ensureModel(id: string) {
     if (mlModel && mlModelId === id) return mlModel
     const status = $("status")
-    const spec = id.startsWith("spec:")
-    const verifierId = spec ? id.slice(5) : id
-    const drafterId = spec ? "ara-diac-layerdrop-1.0-int4" : ""
-    const wanted = spec ? [drafterId, verifierId] : [verifierId]
+    const wanted = [id]
     const models: Record<string, Awaited<ReturnType<typeof imf.IMFModel.fromZipBytes>>> = {}
     for (const wantedId of wanted) {
       status.hidden = false
@@ -204,12 +194,7 @@ import catalogue from "../data/maps-catalogue.json"
       })
       models[wantedId] = await imf.IMFModel.fromZipBytes(resolved.bytes)
     }
-    mlModel = spec
-      ? (new imf.SpeculativeModel(models[drafterId]!, models[verifierId]!) as unknown as {
-          translate(t: string, m?: number): Promise<string>
-          stats(): { accepted: number; drafted: number; blocks: number } | undefined
-        })
-      : (models[verifierId] as unknown as { translate(t: string, m?: number): Promise<string> })
+    mlModel = models[id] as unknown as { translate(t: string, m?: number): Promise<string> }
     mlModelId = id
     status.hidden = true
     return mlModel
@@ -234,15 +219,6 @@ import catalogue from "../data/maps-catalogue.json"
       }
       const result = await worker.transliterate(systemSelect.value, text)
       out.textContent = result
-      if (modelId?.startsWith("spec:") && mlModel) {
-        const stats = (
-          mlModel as { stats?: () => { accepted: number; drafted: number; blocks: number } }
-        ).stats?.()
-        if (stats) {
-          status.hidden = false
-          status.textContent = `Speculative: ${stats.accepted}/${stats.drafted} draft tokens accepted across ${stats.blocks} verifier passes`
-        }
-      }
       const share = new URLSearchParams({
         q: ($("input") as HTMLTextAreaElement).value,
         sys: systemSelect.value,