fix: log non-OpenAI format= assumption once at init (#1502)

[r72ntkttwb0f2i]

Summary

• Move the non-OpenAI format= assumption INFO log from every generate() call to OpenAIBackend.__init__, so local / OpenAI-compatible providers do not spam logs on each request (OpenAIBackend logs a per-call INFO message when using non-OpenAI providers — should log once at init or at DEBUG #1502).
• Add unit tests covering init-once logging for non-OpenAI URLs and no log for the OpenAI platform.

Test plan

• pytest test/backends/test_openai_unit.py::test_non_openai_format_assumption_logged_once_at_init test/backends/test_openai_unit.py::test_openai_platform_skips_format_assumption_log
• Confirm a localhost / OpenAI-compatible backend logs the message once at construction, not per format= call

Fixes #1502

[jakelorocco]

Thanks for the contribution @r72ntkttwb0f2i! I've kicked off the acceptance tests and we should be able to merge after they are done. The change seems reasonable.

[psschwei]

looks like the linters failed:

# Format code
uv run ruff format .

# Lint code
uv run ruff check .

# Fix auto-fixable issues
uv run ruff check --fix .

# Type check
uv run mypy .

[planetf1]

Verified against current HEAD, the CI job logs, and issue #1502's own proposed-fix text (not just re-reading the diff). Two real, pre-existing issues below with one-line fixes; the CI failure across all three Python quality jobs is a plain ruff format miss, not a logic problem, also fixed inline.

One finding falls outside this diff's hunks so GitHub won't let me anchor it inline: mellea/backends/openai.py:189 — self._base_url = base_url never falls back to os.getenv("OPENAI_BASE_URL"). If base_url isn't passed but the env var is set, self._base_url stays None, so _server_type() a few lines down defaults to OPENAI and this PR's new init-time log never fires — even though the SDK will actually route to whatever host the env var names. Pre-existing, but worth closing now that this log is being promoted to the authoritative construction-time notice. Suggested fix: self._base_url = base_url or os.getenv("OPENAI_BASE_URL").

One more thing, not tied to a single line: #1493 (open, targeting #1491) rewrites the same block in _generate_from_chat_context_standard — the if _format is not None: if self._server_type == _ServerType.OPENAI: ... else: ... split — as part of unifying schema handling. Whichever of #1493 / #1514 merges second will hit a source conflict there. Worth sequencing deliberately rather than merging blind.

[planetf1]

Pre-existing typo carried over from the deleted log line: the repo slug elsewhere in the codebase (README, GOVERNANCE, bedrock.py, ollama.py) is always hyphenated. As written this is a dead link.

Suggested change

	"github.com/generative_computing/mellea with the stack trace and "
	"github.com/generative-computing/mellea with the stack trace and "

[planetf1]

Flagging in case another reviewer raises this: moving the log to fire on server type alone, regardless of whether format= is ever used, isn't scope creep — issue #1502's own "Proposed fix, Option 1" asks for exactly this ("tie the message to backend setup rather than the generate loop"). This matches the issue as written.

[planetf1]

This is the actual cause of the quality (3.11/3.12/3.13) CI failures — ruff format collapses these onto one line each. Not a logic issue; same fix needed at line 466-470 below.

Suggested change

	OpenAIBackend(
	model_id="gpt-4o",
	api_key="fake-key",
	base_url="http://localhost:9999/v1",
	)
	OpenAIBackend(
	model_id="gpt-4o",
	api_key="fake-key",
	base_url="http://localhost:9999/v1",
	)
	OpenAIBackend(
	model_id="gpt-4o", api_key="fake-key", base_url="http://localhost:9999/v1"
	)
	OpenAIBackend(
	model_id="gpt-4o", api_key="fake-key", base_url="http://localhost:9999/v1"
	)

[planetf1]

Same ruff format fix as above.

Suggested change

	OpenAIBackend(
	model_id="gpt-4o",
	api_key="fake-key",
	base_url="https://api.openai.com/v1",
	)
	OpenAIBackend(
	model_id="gpt-4o", api_key="fake-key", base_url="https://api.openai.com/v1"
	)

[r72ntkttwb0f2i]

Thanks for the review - addressed in the latest commits:

uff format on the new unit tests

• hyphenated generative-computing in the issue URL
• self._base_url = base_url or os.getenv("OPENAI_BASE_URL") so init-time _server_type / logging matches the host the SDK would use
• small test covering the env-only non-OpenAI case

Re: sequencing with #1493 - happy to rebase if that lands first.