ds4_server: Add /health endpoint that returns HTTP 200 once model is fully loaded

[mcmalayalam]

Summary

Adds a simple /health HTTP endpoint to ds4_server.c, which returns 200 OK once a model is fully loaded, and 503 Service Unavailable otherwise.

References to existing issues and pull requests

This fixes issue #334, so that tools such as llama-swap can start ds4-server and detect when it is ready to receive requests.

I note that there is open PR #326 which implements a more full-featured /health endpoint along with other endpoints. If that PR is favoured and merged I'm more than happy to close this one.

Test evidence

Manually tested with llama-swap and ds4-server on Strix Halo.

ds4-server logs:

ds4: ROCm startup model preparation covered 80.76 GiB of tensor spans in 0.642s
ds4: ROCm preparing model tensor mappings
ds4: ROCm startup model preparation covered 3.55 GiB of tensor spans in 0.012s
ds4: rocm backend initialized for graph diagnostics
0609 21:46:11 ds4-server: context buffers 4460.31 MiB (ctx=131072, backend=rocm, prefill_chunk=8192, raw_kv_rows=8192, compressed_kv_rows=32770)
0609 21:46:12 ds4-server: KV disk cache /tmp/ds4-kv (budget=8192 MiB, cross-quant=accept, min=512, cold_max=30000, continued=10000, trim=32, align=2048, hit_half_life=21600s)
0609 21:46:12 ds4-server: listening on http://127.0.0.1:5816

llama-swap logs:

2026-06-09T21:46:12+01:00 [INFO] <deepseek-v4-flash-ds4> Health check passed on http://localhost:5816/health