feat(server): add legacy /v1/completions endpoint for the inbuilt inference server

[akushonkamen]

Fixes #170

Problem

The built-in inference server only exposes the chat protocol at /v1/chat/completions (optillm/server.py:759). OpenAI-legacy clients that speak the text completions protocol (client.completions.create, older SDKs/tools) cannot use optillm's local inference.

What this PR does

Extracts the request_config parsing, approach/model resolution, client building and dispatch of proxy() into a shared _execute_completion(data, messages, bearer_token) pipeline (CompletionOutcome / CompletionError), then adds POST /v1/completions on top of it — no HTTP self-call, both endpoints share one dispatch path.

First-version scope (requesting confirmation from @codelion):

• prompt (str or list[str]) maps to a single user message; the pipeline runs once per prompt and n still expands attempts, so choices accumulate (n completions per prompt, OpenAI legacy semantics).
• Response shape: {"id": "cmpl-...", "object": "text_completion", "created": <unix>, "model": <dispatch model>, "choices": [{"text", "index", "logprobs": null, "finish_reason": "stop"}], "usage": {completion_tokens, completion_tokens_details.reasoning_tokens}}.
• stream=true → 400 with a pointer to /v1/chat/completions (first version; happy to add SSE in a follow-up).
• optillm_approach (body key or slug- model prefix), max_tokens/max_completion_tokens, temperature, stop, etc. pass through exactly like the chat endpoint; unsupported legacy-only params (echo, suffix, logprobs=true) are not implemented yet.
• Passthrough completion dicts (the none approach, full-response plugins like proxy, batch mode) are mapped choice-by-choice into text choices; batch mode responses are likewise converted.
• /v1/chat/completions responses are byte-identical to before; conversation logging, batch-mode aggregation and error mapping (500 {"error": ...}) behave as before. Auth is unchanged (the existing @app.before_request API-key check covers the new route automatically).
• README: documents the new endpoint with a curl example in the "Local inference server" section.

Red → green evidence

New offline tests tests/test_completions_endpoint.py (8 tests, Flask test_client() + mocked chat_with_mcts/get_config, following tests/test_approaches.py:335):

• Before the change (unmodified server.py on origin/main): 8 failed — the route does not exist (404s).
• After the change: 8 passed (shape, multi-prompt, n-expansion, stream 400, prompt validation, error 500, chat shape regression guard).

8 failed   (pre-fix:  python -m pytest tests/test_completions_endpoint.py)
8 passed   (post-fix: python -m pytest tests/test_completions_endpoint.py)

Regression tests (all green)

• pytest tests/test_approaches.py — 6 passed (exercises the refactored proxy() via test_client, incl. the issue Bug: Race condition — concurrent requests corrupt MCTS parameters via shared global state #304 request-scoped mcts params test)
• pytest tests/test_plugins.py — 22 passed
• pytest tests/test_proxy_plugin.py — 9 passed
• pytest tests/test_reasoning_simple.py tests/test_ssl_config.py — 30 passed
• OPTILLM_API_KEY=optillm python tests/test_ci_quick.py — all checks passed
• Note: pytest tests/test_batching.py fails at collection with ModuleNotFoundError: No module named 'test_utils' on unmodified origin/main too (pre-existing, unrelated to this PR).
• Lint: repo has no lint/format config (no ruff/black/flake8/mypy); changed files pass python -m compileall and follow existing code style.

Test plan

• python -m pytest tests/test_completions_endpoint.py -v
• python -m pytest tests/test_approaches.py tests/test_plugins.py tests/test_proxy_plugin.py
• OPTILLM_API_KEY=optillm python tests/test_ci_quick.py

---

Transparency note: this PR was prepared with the assistance of an AI coding agent (ZCode); all code changes were reviewed and the full test matrix executed by the submitting author.