fix: harden provider recommendation safety

src/utils/providerRecommendation.test.ts

@@ -83,6 +83,19 @@ test('non-chat embedding models are heavily demoted', () => {
   assert.equal(ranked[0]?.name, 'mistral:7b-instruct')
 })
 
+test('auto-pick ignores non-chat ollama models', () => {
+  const recommended = recommendOllamaModel(
+    [
+      model('nomic-embed-text', { parameterSize: '0.5B' }),
+      model('bge-reranker-v2', { parameterSize: '1.5B' }),
+      model('whisper-large-v3', { parameterSize: '1.6B' }),
+    ],
+    'balanced',
+  )
+
+  assert.equal(recommended, null)
+})
+
 test('benchmark latency can reorder close recommendations', () => {
   const ranked = rankOllamaModels(
     [
@@ -111,6 +124,69 @@ test('benchmark latency can reorder close recommendations', () => {
   assert.equal(benchmarked[0]?.benchmarkMs, 350)
 })
 
+test('unbenchmarked models stay behind benchmarked candidates', () => {
+  const ranked = rankOllamaModels(
+    [
+      model('phi4-mini:4b', {
+        parameterSize: '4B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+      model('mistral:7b-instruct', {
+        parameterSize: '7B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+      model('llama3.1:8b', {
+        parameterSize: '8B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+      model('qwen2.5:14b', {
+        parameterSize: '14B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+    ],
+    'latency',
+  )
+
+  const benchmarked = applyBenchmarkLatency(
+    ranked,
+    {
+      'phi4-mini:4b': 2400,
+      'mistral:7b-instruct': 2200,
+      'llama3.1:8b': 2100,
+    },
+    'latency',
+  )
+
+  assert.ok(benchmarked.slice(0, 3).every(item => item.benchmarkMs !== null))
+  assert.equal(benchmarked[3]?.name, 'qwen2.5:14b')
+  assert.equal(benchmarked[3]?.benchmarkMs, null)
+})
+
+test('coding goal recognizes codestral and devstral families', () => {
+  const ranked = rankOllamaModels(
+    [
+      model('mistral:7b-instruct', {
+        parameterSize: '7B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+      model('codestral:22b', {
+        parameterSize: '22B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+      model('devstral:24b', {
+        parameterSize: '24B',
+        quantizationLevel: 'Q4_K_M',
+      }),
+    ],
+    'coding',
+  )
+
+  assert.deepEqual(ranked.slice(0, 2).map(item => item.name), [
+    'devstral:24b',
+    'codestral:22b',
+  ])
+})
+
 test('goal defaults choose sensible openai models', () => {
   assert.equal(getGoalDefaultOpenAIModel('latency'), 'gpt-4o-mini')
   assert.equal(getGoalDefaultOpenAIModel('balanced'), 'gpt-4o')