test: shrink Test 38 prompt to fit Gemma-4 prefill on CI runner

The 6000-number prompt tokenizes to ~42k tokens on Gemma-4. Its prefill
materializes full attention in one Metal buffer (28 GB), above the CI
runner's 3.5 GB max buffer size, so the server crashed and the stream
never reached [DONE]. 1000 numbers (~7k tokens) stays well under the
limit; the ': connected' check still catches the original regression.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

Author: CodeAndCanvas728

tests/test-server.sh

@@ -1161,9 +1161,12 @@ rm -f /tmp/mlx_inflight_models.json /tmp/mlx_inflight_health.json /tmp/mlx_infli
 # whole prefill inside container.perform before returning headers, so clients
 # saw no bytes for the entire prefill (Bun fetch 10s idle → ECONNRESET/retry).
 # A large prompt makes prefill multi-second; TTFB must stay well under that.
+# ponytail: ~7k tokens. Gemma-4 prefill allocates O(n²) attention in one Metal
+# buffer; the CI runner caps buffers at 3.5 GB (~15k tokens). 42k tokens
+# crashed the server with a 28 GB malloc.
 log "Test 38: streaming TTFB — ': connected' arrives before model prefill"
 
-python3 -c "print(' '.join(f'{i:06d}' for i in range(6000)))" > /tmp/mlx_ttfb_prompt.txt
+python3 -c "print(' '.join(f'{i:06d}' for i in range(1000)))" > /tmp/mlx_ttfb_prompt.txt
 jq -nc --arg m "$MODEL" --rawfile p /tmp/mlx_ttfb_prompt.txt \
     '{model:$m, stream:true, max_tokens:5, messages:[{role:"user",content:("Summarize this list in one word:\n"+$p)}]}' \
     > /tmp/mlx_ttfb_body.json