ci(local-model): hold a row that names its runner to the build a Mac pulls

The catalogue that names each row's runner (feat/local-model-catalog-v2: `runner` is `mlx` or
`llamacpp`, and a row's memory figures follow that runner's rule) needs the weekly check to
notice when a tag starts serving a Mac another runner: a tag that gains an mlx child, or a
list that loses it. The size check usually catches that, but not when the two builds come out
about the same size, and the figures are wrong either way.

classifyModel() now raises a `runner` finding when the row names a runner and the build a Mac
selects runs on another one. catalogueRunner() maps the registry's names into the catalogue's:
`mlx` stays `mlx`; `llamacpp`, and `ggml` (what Ollama calls a GGUF manifest that names no
runner), are `llamacpp`, because Ollama 0.40.15 serves every GGUF build through llama-server.
A catalogue without the field (the one on this branch today) is never held to a runner.

Run live on 2026-11-08 against the v2 catalogue's eight rows: exit 0, every runner matches.
Tests: the mapping, the finding (both directions, ggml, an unknown runner, a row without the
field), the report text, and runCheck against the fake registry with a plain GGUF manifest.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

scripts/check-local-model-catalog.mjs

@@ -23,7 +23,10 @@
  * which is NOT what a Mac pulls: that is how `qwen3.6:35b-a3b` came to be "2% above its
  * manifest" in the catalogue header. So this check asks the way Ollama asks, picks the child
  * the way a Mac picks it, and compares the row's `bytes` with that child's pull total (every
- * layer plus the config, which is exactly the `total` a real pull reported on a Mac).
+ * layer plus the config, which is exactly the `total` a real pull reported on a Mac). A row
+ * that names its `runner` is also held to that child's runner (`catalogueRunner()`), because
+ * the memory figures of a row follow its runner and a change of runner can come with almost
+ * the same size.
  *
  * WHAT IS FETCHED, and nothing else:
  *  - `GET /v2/library/<name>/manifests/<tag>`, the tag's manifest or manifest list.
@@ -195,6 +198,18 @@ export function metadataForConfig(config) {
   return { runner: '', format };
 }
 
+/**
+ * A runner name in the catalogue's words (`LocalModelEntry.runner`): `mlx`, or `llamacpp` for a
+ * GGUF build, which Ollama 0.40.15 serves through llama-server whether the manifest names it
+ * `llamacpp` or names nothing (`ggml`, see `metadataForConfig()`). Undefined for anything else.
+ */
+export function catalogueRunner(runner) {
+  const name = canonicalRunner(runner);
+  if (name === 'mlx') return 'mlx';
+  if (name === 'llamacpp' || name === 'ggml') return 'llamacpp';
+  return undefined;
+}
+
 /** `[major, minor, patch]` of a version or `v`-tag, or undefined when it is not one. */
 export function parseVersion(text) {
   const match = typeof text === 'string' ? VERSION_PATTERN.exec(text.trim()) : null;
@@ -221,11 +236,14 @@ export function withinTolerance(expected, actual, tolerance = SIZE_TOLERANCE_FRA
 }
 
 /**
- * The findings for one catalogue row. `row` is `{ id, bytes, digest? }` (`digest` only once the
- * catalogue stores one); `observed` is what `inspectModel()` returned for it. A row may have
- * several findings; none means it is clean.
+ * The findings for one catalogue row. `row` is `{ id, bytes, runner?, digest? }` (`runner` once
+ * the catalogue names it, `digest` only once the catalogue stores one); `observed` is what
+ * `inspectModel()` returned for it. A row may have several findings; none means it is clean.
  *
  *  - `gone`: the tag is not in the registry, so a pull fails.
+ *  - `runner`: a Mac now pulls a build of another runner than the row names (a tag that became
+ *    a list with an mlx child, or the other way round). The two runners spend memory
+ *    differently, so the row's memory figures no longer describe what runs, whatever the size.
  *  - `size`: what a Mac downloads is outside the tolerance of the row's `bytes`.
  *  - `digest`: the row pins a digest and a Mac would now pull another one.
  *  - `requires`: the build a Mac pulls declares a newer Ollama than the pinned runtime, so the
@@ -235,6 +253,10 @@ export function classifyModel(row, observed, pinnedVersion) {
   if (observed.gone) return [{ kind: 'gone' }];
   const findings = [];
   const mac = observed.mac;
+  if (row.runner !== undefined) {
+    const actual = catalogueRunner(mac.runner) ?? 'unknown';
+    if (actual !== row.runner) findings.push({ kind: 'runner', expected: String(row.runner), actual });
+  }
   if (!withinTolerance(row.bytes, mac.bytes)) {
     findings.push({ kind: 'size', expected: row.bytes, actual: mac.bytes, fraction: sizeDifference(row.bytes, mac.bytes) });
   }
@@ -314,7 +336,13 @@ function buildName(variant) {
 function modelResult(model) {
   if (model.error) return 'check failed';
   if (model.findings.length === 0) return 'ok';
-  const names = { gone: 'tag gone', size: 'size drift', digest: 'digest changed', requires: 'needs a newer runtime' };
+  const names = {
+    gone: 'tag gone',
+    runner: 'runner changed',
+    size: 'size drift',
+    digest: 'digest changed',
+    requires: 'needs a newer runtime',
+  };
   return model.findings.map((finding) => names[finding.kind]).join(', ');
 }
 
@@ -323,6 +351,13 @@ function describeModelFinding(model, finding) {
   switch (finding.kind) {
     case 'gone':
       return `${id}: the registry no longer has this tag, so the download fails. Replace or remove the row.`;
+    case 'runner':
+      return (
+        `${id}: a Mac now pulls the ${buildName(model.observed.mac)} build, which runs on ` +
+        `\`${finding.actual}\`, while the row's \`runner\` says \`${displayName(finding.expected, ['mlx', 'llamacpp'])}\`. ` +
+        'The two runners spend memory differently, so the row\'s memory figures no longer describe what runs: ' +
+        'recompute them for the new runner, or pin the row to a single-build tag.'
+      );
     case 'size':
       return (
         `${id}: a Mac downloads the ${buildName(model.observed.mac)} build, ${groupDigits(finding.actual)} bytes, ` +
@@ -423,7 +458,7 @@ export function renderReport(result) {
     if (model.error) details.push(`- \`${model.id}\`: check failed: ${oneLine(model.error)}`);
     for (const finding of model.findings) details.push(`- ${describeModelFinding(model, finding)}`);
     const variants = model.observed?.variants ?? [];
-    if (model.findings.some((finding) => finding.kind === 'size') && variants.length > 1) {
+    if (model.findings.some((finding) => finding.kind === 'size' || finding.kind === 'runner') && variants.length > 1) {
       const offered = variants
         .map((variant) => `${buildName(variant)} ${groupDigits(variant.bytes)}${variant === model.observed.mac ? ' (a Mac takes this one)' : ''}`)
         .join('; ');
@@ -695,12 +730,14 @@ export async function runCheck({ catalog, pin, fetchFn, now, githubToken }) {
   // Rows in parallel, each row's requests one after another: a handful of rows, a few requests each.
   result.models = await Promise.all(
     catalog.models.map(async (row) => {
-      // `digest` is not in the catalogue schema yet: see the file header.
+      // `digest` is not in the catalogue schema yet: see the file header. `runner` is in the
+      // schema from the catalogue that names each row's runner on; before that it is absent.
       const digest = /** @type {{ digest?: unknown }} */ (row).digest;
+      const runner = /** @type {{ runner?: unknown }} */ (row).runner;
       const entry = { id: row.id, bytes: row.bytes, figures: row.figures, digest, findings: [] };
       try {
         entry.observed = await inspectModel(fetchFn, row.id);
-        entry.findings = classifyModel({ id: row.id, bytes: row.bytes, digest }, entry.observed, pin.version);
+        entry.findings = classifyModel({ id: row.id, bytes: row.bytes, runner, digest }, entry.observed, pin.version);
       } catch (err) {
         entry.error = err instanceof CheckFailure ? err.message : describeError(err);
       }