fix(dashboard): resolve combo limits with the server's capabilities (#4360)

The combos page computed combo capabilities in the browser using
aggregateComboCapabilities, falling back to pattern defaults for models
without exact entries because the synced model catalog is server-only.

Allow callers to supply a resolveCaps callback (e.g. from useModelCaps)
merged over the local tables, preserving non-limit capability flags.

tests/unit/combo-caps-resolver.test.js

@@ -0,0 +1,67 @@
+import { describe, expect, it } from "vitest";
+
+import { aggregateComboCapabilities, getCapabilitiesForModel } from "../../open-sse/providers/capabilities.js";
+
+// A combo's limits are the conservative aggregate of its members: ctx = min,
+// maxOutput = max. Resolving those members needs the synced model catalog, which
+// is server-only (it reads a file), so the browser bundle falls back to the
+// generic patterns. The dashboard computed its badges there and under-reported:
+// /v1/models and pi-settings (both server-side) said 1M while the badge said 200k.
+//
+// resolveCaps lets a caller hand in the server's answer. It must only override
+// what it carries — the local tables still own tools/pdf/audio/video/thinking*.
+const GLM53_FED = { vision: true, search: false, reasoning: true, contextWindow: 1_000_000, maxOutput: 131_072 };
+
+describe("aggregateComboCapabilities: resolveCaps override", () => {
+  const models = ["glm-cn/glm-5.3", "deepseek-v4.1-flash"];
+
+  it("falls back to the pattern default without a resolver", () => {
+    const caps = aggregateComboCapabilities(models);
+    // glm-5.3 has no exact entry, so the *glm-5.3* pattern gives 200k and caps the combo.
+    expect(caps.contextWindow).toBe(200_000);
+  });
+
+  it("uses the fed limits when a resolver supplies them", () => {
+    const resolver = (fullId) => (fullId === "glm-cn/glm-5.3" ? GLM53_FED : null);
+    const caps = aggregateComboCapabilities(models, null, resolver);
+    expect(caps.contextWindow).toBe(1_000_000);
+  });
+
+  it("keeps the fields the override does not carry", () => {
+    const plain = aggregateComboCapabilities(models);
+    const fed = aggregateComboCapabilities(models, null, (id) => (id === "glm-cn/glm-5.3" ? GLM53_FED : null));
+    // The override carries no tools/pdf/thinking fields, so those must be unchanged.
+    for (const field of ["tools", "pdf", "audioInput", "videoInput", "imageOutput", "audioOutput", "thinkingFormat"]) {
+      expect(fed[field]).toEqual(plain[field]);
+    }
+  });
+
+  it("still applies the conservative rule across members", () => {
+    const resolver = (fullId) => (fullId === "glm-cn/glm-5.3" ? GLM53_FED : null);
+    const caps = aggregateComboCapabilities(models, null, resolver);
+    // Only glm-5.3 was fed 1M; deepseek-v4.1-flash resolves locally to 1M, so min stays 1M.
+    // Feeding a *smaller* value for one member must pull the aggregate down.
+    const smaller = aggregateComboCapabilities(models, null, (id) => (id === "glm-cn/glm-5.3" ? { ...GLM53_FED, contextWindow: 64_000 } : null));
+    expect(smaller.contextWindow).toBe(64_000);
+    expect(caps.maxOutput).toBe(384_000); // max across members, from deepseek
+  });
+
+  it("passes the resolver into nested combos", () => {
+    const lookup = {
+      zap: ["deepseek-v4.1-flash", "glm-cn/glm-5.3-flash"],
+      "deepseek-v4.1-flash": ["cmc/deepseek/deepseek-v4.1-flash", "ocg/deepseek-v4.1-flash"],
+    };
+    const seen = [];
+    const resolver = (fullId) => { seen.push(fullId); return fullId === "glm-cn/glm-5.3-flash" ? { contextWindow: 1_000_000 } : null; };
+    aggregateComboCapabilities(lookup.zap, lookup, resolver);
+    // The nested combo's own members were resolved with the same resolver.
+    expect(seen).toContain("cmc/deepseek/deepseek-v4.1-flash");
+    expect(seen).toContain("ocg/deepseek-v4.1-flash");
+  });
+
+  it("leaves the plain two-argument call unchanged", () => {
+    const caps = aggregateComboCapabilities(["kimi/kimi-k3"], null);
+    expect(caps).toEqual(aggregateComboCapabilities(["kimi/kimi-k3"]));
+    expect(caps.contextWindow).toBe(getCapabilitiesForModel("kimi", "kimi-k3").contextWindow);
+  });
+});