Expose a loaded LlamaModel over an OpenAI-compatible HTTP API so editors that
speak the OpenAI protocol (e.g. VS Code Copilot "Custom Endpoint") can drive a
local GGUF model in-process, with no extra runtime dependency (the JDK's built-in
com.sun.net.httpserver).

Native (jllama.cpp, json_helpers.hpp):
- requestChatCompletionStream + receiveChatCompletionChunk: streaming chat that
  formats each result as an OpenAI chat.completion.chunk (incl. streamed
  delta.tool_calls) via TASK_RESPONSE_TYPE_OAI_CHAT. Uniform {data,stop} envelope
  built by the new pure helper wrap_stream_chunk (+5 C++ tests; 445 total).

Java:
- LlamaModel.streamChatCompletion(params, sink) drives the native chunk stream;
  ChatStreamChunkParser parses the envelope.
- New net.ladenthin.llama.server package: OpenAiCompatServer (POST
  /v1/chat/completions streaming via SSE + non-streaming, GET /v1/models, SSE
  heartbeats to survive CPU prefill, optional bearer auth, binds 127.0.0.1),
  OpenAiServerConfig, OpenAiRequestMapper (messages/tools forwarded verbatim),
  OpenAiSseFormatter, plus a ChatBackend seam for model-free HTTP tests and a
  CLI launcher.
- module-info: export the server package, requires jdk.httpserver.

Tests: pure-Java mapper/SSE/parser unit tests + a full HTTP test driven over a
socket with a fake backend (no model/native lib). Architecture rule refined to
allow the supported com.sun.net.httpserver API. Runnable OpenAiServerExample and
README/CLAUDE/TODO documentation.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ

…273)

OpenAiCompatServer's CLI main() called Integer.parseInt on --ctx, --gpu-layers,
--parallel and --port without guarding NumberFormatException, so a non-numeric
value (e.g. "--port abc") crashed with a raw stack trace — flagged by CodeQL as
"missing catch of NumberFormatException" (5 alerts).

Consolidate the numeric parsing into one try/catch that prints a clear message and
returns (no System.exit — the noSystemExit architecture rule forbids it; this
mirrors the existing missing-"--model" usage path), and parse --ctx once instead
of twice.

Verified: compiles clean under -Werror / NullAway / Checker; LlamaArchitectureTest
(noSystemExit) passes.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ

…sort

The clang-format enforcement commit ran with SortIncludes: CaseSensitive, which
alphabetically reordered includes and moved jni_helpers.hpp / json_helpers.hpp
ahead of the server-*.h headers that define the `json` alias they depend on. CI
then failed to build jllama with "'json' does not name a type" cascading through
json_helpers.hpp / jni_helpers.hpp / jllama.cpp on the manylinux, Linux aarch64,
Android and C++-test compilers. (A local build masks it — the local toolchain
resolves `json` regardless of include order.)

Set SortIncludes: Never in .clang-format (the project has order-sensitive includes,
documented in CLAUDE.md) and restore the required order — server-*.h + utils.hpp
before the helper headers — in jllama.cpp and the affected C++ test files. Document
the constraint in CLAUDE.md. clang-format --dry-run --Werror stays clean; the
affected targets rebuild and the C++ tests pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ

The model-download steps used `curl -L --fail` with no retry, so a single
transient HTTP 429 (HuggingFace rate-limiting) failed the whole job — e.g. the
default codellama-7b.Q2_K.gguf download in run 27778671637. Add
`--retry 5 --retry-all-errors` to all 31 model-download curls (the Linux bash
blocks and the Windows pwsh block). curl honors the server's Retry-After header
and backs off on 429/5xx/connection errors; `--fail` still fails the step only
after the retries are exhausted.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ

The Win32 (x86) C++ test job intermittently failed at build-time
gtest_discover_tests. llama/ggml/mtmd are linked statically into one
large jllama_test binary; on 32-bit Windows its startup plus
--gtest_list_tests enumeration sits near the default 5s discovery
timeout on shared CI runners. The same b9682 binary discovered within
5s in the #239 merge run but was killed at the 5s timeout in this run
(process still alive, empty output — a timeout, not a crash); the b9682
upgrade and 5 newly added tests nudged a marginal case over the limit.
x64, Linux and macOS finish well under the default and are unaffected.

Raise DISCOVERY_TIMEOUT to 120s (a maximum, so fast platforms still
return immediately), which keeps full C++ test coverage on x86 rather
than skipping the binary there. Verified locally: 445/445 C++ tests
still pass.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ

Make PR #240 mergeable on top of the NanoHTTPD OpenAI server that landed
via #242. Both OpenAI-compatible server implementations now coexist in
net.ladenthin.llama.server, pending a "best of both" consolidation
(tracked in TODO.md):

- OpenAiCompatServer (PR #240): JDK com.sun.net.httpserver, streaming SSE
  with delta.tool_calls, no new runtime dependency.
- LlamaServer (#242): NanoHTTPD, non-streaming, fat-jar Main-Class, plus
  /v1/completions, /v1/embeddings and /health.

Conflicts resolved:
- server/package-info.java (add/add): documents both servers + pending merge.
- README.md, CLAUDE.md: keep both server sections under one heading.
- TODO.md: add a consolidation task; note SSE is already solved by
  OpenAiCompatServer and that com.sun.net.httpserver is the supported
  jdk.httpserver module, not an internal com.sun.. API.

Auto-merged and verified consistent: LlamaModel.java (distinct native
methods on each side), publish.yml, pom.xml (NanoHTTPD + assembly),
spotbugs-exclude.xml, and LlamaArchitectureTest.java — main's Server
layer already permits this session's server classes (they touch only the
Api root + Parameters), and the noInternalJdkImports com.sun.net.httpserver
exception merges alongside.

Verified: mvn compile + test-compile clean; 64 model-free tests pass
(LlamaArchitectureTest + both servers' unit/HTTP tests, integration test
self-skips without a model); javadoc jar builds clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014L2dLbAtwdq7C6a2gFRsQQ