fix: gracefully handle Ollama models without tool support

Check model capabilities via /api/show before sending tools.
Models without 'tools' capability get requests without tools
(they can still answer, just without tool use). Result is cached
per client instance. Defense-in-depth: 'does not support' added
to retry nonRetryablePatterns to avoid wasting retries on
permanent errors.

src/models/retry.test.ts

@@ -38,6 +38,11 @@ describe('isRetryable', () => {
     expect(isRetryable(error, DEFAULT_RETRY_CONFIG.nonRetryablePatterns)).toBe(false);
   });
 
+  it('returns false for "does not support" errors', () => {
+    const error = new Error('registry.ollama.ai/library/qwen2.5:latest does not support tools');
+    expect(isRetryable(error, DEFAULT_RETRY_CONFIG.nonRetryablePatterns)).toBe(false);
+  });
+
   it('is case-insensitive when matching patterns', () => {
     const error = new Error('AUTHENTICATION error');
     expect(isRetryable(error, DEFAULT_RETRY_CONFIG.nonRetryablePatterns)).toBe(false);
@@ -102,6 +107,15 @@ describe('withRetry', () => {
     expect(fn).toHaveBeenCalledTimes(1);
   });
 
+  it('does not retry "does not support tools" errors', async () => {
+    const fn = vi.fn().mockRejectedValue(
+      new Error('registry.ollama.ai/library/qwen2.5:latest does not support tools'),
+    );
+
+    await expect(withRetry(fn, fastConfig)).rejects.toThrow('does not support tools');
+    expect(fn).toHaveBeenCalledTimes(1);
+  });
+
   it('converts non-Error throws to Error objects', async () => {
     const fn = vi.fn().mockRejectedValue('string error');