feat: add protocol= + base_url= for compatible APIs

[Kamilbenkirane]

Summary

• Add protocol= + base_url= parameters to create_client() and domain namespaces for pointing celeste at any compatible API (MiniMax, vLLM, Ollama, LocalAI, etc.)
• Separate provider and protocol as independent fields on ModalityClient — a protocol is NOT a provider
• Wire base_url through to _build_url() as a client instance attribute
• BYOA auth for protocol path (defaults to NoAuth, accepts api_key= or auth=)
• Remove dead base_url plumbing from per-call method signatures

Usage

# OpenAI-compatible (openresponses protocol)
await celeste.text.generate("Hello", model="lfm2:latest", protocol="openresponses", base_url="http://localhost:11434")

# OpenAI Chat Completions-compatible (e.g. MiniMax, vLLM)
await celeste.text.generate("Hello", model="MiniMax-M2.5", protocol="chatcompletions", base_url="https://api.minimax.io", api_key="XX-YOUR-API-KEY")

# Provider with custom base URL
client = create_client(provider="openai", base_url="https://my-proxy.com", model="gpt-4o", modality="text")

Design decisions

• protocol replaces provider for the compatible API case — same concept (wire format identity) without naming a company
• ModalityClient.provider: Provider | None — None for protocol-only clients
• ModalityClient.protocol: Protocol | None — None for provider-only clients
• Model.provider: Provider | None — None for protocol-path models (standard Python Optional pattern)
• base_url without protocol defaults to openresponses (most common compatible format)

Test plan

• 527 unit tests pass
• 15 new tests for protocol + base_url paths
• mypy passes (0 errors, no type:ignore on new code)
• Smoke tested with local Ollama (lfm2:latest via protocol="openresponses")
• Smoke tested provider path (provider="ollama") still works

Closes #230
Closes #237

Co-Authored-By: Xinyue 7869833+XinyueZ@users.noreply.github.com

[claude]

Code review

Bug: provider is dropped when provider= is not passed explicitly (src/celeste/init.py:276-283)

When create_client() is called without an explicit provider= argument (the common case — e.g. create_client(model="gpt-4o", modality=Modality.TEXT)), resolved_provider is None. The client is then constructed with both provider=None and protocol=None. When any HTTP call is made, ModalityClient.http_client (client.py:175-181) raises ClientNotFoundError because it checks self.provider first, then self.protocol, and both are None.

Before this PR the line read provider=resolved_model.provider, which always carried the provider derived from model resolution. The change to provider=resolved_provider breaks the standard provider path any time the caller omits provider=.

Fix: change line 279 to:

        provider=resolved_provider or resolved_model.provider,