fix: AI SDK compatibility, rootless Docker, preflight checks

[aWN4Y25pa2EK]

Summary

• Add --reasoning-format none to Qwen3.5-9B profiles to prevent AI SDK crash on empty <think> blocks
• Fix docker-compose.yml for rootless Docker (remove failing sysctls/ulimits, fix prometheus duplicate volumes)
• Add startup preflight check that warns about suboptimal host kernel params (BBR, NUMA, swappiness, GPU persistence)
• Update AGENTS.md OpenCode config to working @ai-sdk/openai-compatible setup
• Add comprehensive TROUBLESHOOTING.md

Problem

The Unsloth Qwen3.5-9B GGUF always emits <think> tokens even with thinking disabled. When llama-server's default --reasoning-format auto detects these, it produces a reasoning_content field in the API response. The Vercel AI SDK's extractReasoningMiddleware crashes on empty reasoning blocks (vercel/ai #12054), causing:

• @ai-sdk/openai: "text part msg_... not found" crash
• @ai-sdk/openai-compatible: Raw <tool_call> XML in content instead of executing tools

Additionally, docker-compose.yml had container-level sysctls (tcp_congestion_control=bbr, busy_read, busy_poll) and ulimits.memlock that fail on rootless Docker. These settings are already applied at the host level by scripts/host-setup.sh, but there was no warning when they weren't applied.

Solution

Server-side: --reasoning-format none in PROFILE_EXTRA_ARGS tells llama-server to leave <think> tags as plain text in content and never produce a reasoning_content field.

Client-side: Use @ai-sdk/openai-compatible (not @ai-sdk/openai) which doesn't have the extractReasoningMiddleware.

Docker: Remove rootless-incompatible settings from compose, add comment pointing to host-setup.sh.

Preflight check: New preflight_check() in the entrypoint reads /proc/sys/ on startup and warns about missing host tuning:

[foundry] Checking host kernel parameters...
[foundry]   tcp_congestion_control = cubic (expected: bbr) -- BBR reduces token streaming latency
[foundry]
[foundry]   1 performance issue(s) detected. Run on the host:
[foundry]     sudo ./scripts/host-setup.sh

Files Changed

File	Change
models/qwen3.5-9b/profiles/rtx5090.sh	Add --reasoning-format none with comment
models/qwen3.5-9b/profiles/default.sh	Add --reasoning-format none
scripts/entrypoint.sh	Add preflight_check() for host kernel param warnings
models/qwen3.5-9b/entrypoint.sh	Sync with scripts/entrypoint.sh
docker-compose.yml	Remove rootless-incompatible sysctls/ulimits, fix prometheus, add host-setup.sh comment
AGENTS.md	Update OpenCode config, add troubleshooting cross-refs
TROUBLESHOOTING.md	New comprehensive troubleshooting guide
.gitignore	Add /opencode.json, remote-build.log

Testing

Verified on eae@192.168.0.219 (RTX 5090, rootless Docker):

• Preflight check correctly detects missing BBR
• --reasoning-format none confirmed: no reasoning_content in API responses
• Streaming tool calls work with @ai-sdk/openai-compatible
• 4 parallel slots, 262K context/slot, all 33/33 layers on GPU