fix: accept mapping venice_parameters in chat.completions type hints (#58)

[sethbang]

Closes #58.

The problem

Both chat.completions.create() overloads annotated venice_parameters: VeniceParameters | None, which is narrower than what the code accepts. A mapping of the same fields has always been validated and coerced by ChatCompletionRequest, but type checkers rejected the call:

error: No overload variant of "create" of "ChatCompletions" matches argument
types "str", "list[dict[str, str]]", "dict[str, str]"  [call-overload]

Because create() is overloaded, the diagnostic doesn't name the offending argument — it reports the whole call as unmatched, which is harder to act on than the messages= version was.

Reported in #1 alongside the messages= variant. That half shipped in 2.0.1; this one didn't.

The fix

• Mapping added to the collections.abc import — it wasn't there.
• Both create() overloads widened to VeniceParameters | Mapping[str, Any] | None.
• The venice_parameters: docstring extended, including the extra="allow" note.

ChatCompletionRequest.venice_parameters deliberately stays model-only — same boundary decision as messages=: that field is where coercion happens.

Annotation-only. Unlike messages=, nothing reads venice_parameters attributes before the request model — it is popped from kwargs and passed straight through — so no runtime normalization was needed. resources/responses.py already annotates it Any | None and was never affected.

Verification

Every unit test here would have passed before this change, since runtime always accepted dicts. They pin the equivalence the annotation now advertises; the type-checker probe is what actually discriminates, so it was run in both directions:

probe	before	after
venice_parameters={"enable_web_search": "on"}	[call-overload]	Success
venice_parameters=VeniceParameters(...)	Success	Success
venice_parameters=42	error	still errors

The last row is the one that matters — widening to Mapping[str, Any] must not quietly turn the hint into Any.

• mypy src/ — clean, 172 source files
• ruff check + ruff format --check — clean
• 12 new tests in tests/unit/resources/test_chat_venice_parameters_mapping.py: request-model coercion, body equivalence through create(), extra="allow" parity between the two forms, invalid values still raising before any request, and E2EE engagement via a mapping (venice_parameters={"enable_e2ee": True} must engage the encrypted flow, or it would silently send plaintext)
• tests/unit/resources/ — 736 passed

make test not run — that one's yours.

Accepting strictly more than before, so this is non-breaking → patch release.

🤖 Generated with Claude Code