feat(providers): add DeepInfra provider and curated models (#619)

[cagdasyurekli]

Summary

This PR addresses and closes #619 by registering DeepInfra as a first-class model provider and curating its latest generation of frontier open-weight models (DeepSeek V4, GLM 5.3, Kimi K3) with verified context window metadata.

Motivation & Background

DeepInfra hosts high-performance inference endpoints for leading open-weight frontier models. While users could previously configure a generic custom endpoint, doing so caused:

1. The desktop UI's context fill meter to show "Context meter unavailable for custom models".
2. The context compaction engine (TurnEngine._compaction_config) to fall back to default thresholds, prematurely compacting long-context models like GLM 5.3 and Kimi K3 (which support 1,000,000 tokens).

Changes

1. Provider Registration (coworker/providers/registry.py):

   • Registered deepinfra in DESCRIPTORS via _compat() pointing to https://api.deepinfra.com/v1/openai.
   • Default environment variable: DEEPINFRA_API_KEY.
   • Recommended model: deepseek-ai/DeepSeek-V4-Flash.
   • Pre-filled editable endpoint with help text.

2. Curated Model Matrix (coworker/providers/matrix.py):

   • Added 8 curated entries with verified context windows and capabilities:
     • deepinfra:deepseek-ai/DeepSeek-V4-Flash: 128,000 tokens (_AGENTIC)
     • deepinfra:deepseek-ai/DeepSeek-V4-Flash-0731: 128,000 tokens (_AGENTIC)
     • deepinfra:deepseek-ai/DeepSeek-V4-Pro: 128,000 tokens (_AGENTIC)
     • deepinfra:deepseek-ai/DeepSeek-V4-Pro-0731: 128,000 tokens (_AGENTIC)
     • deepinfra:zai-org/GLM-5.3-Flash: 1,000,000 tokens (_AGENTIC_VISION)
     • deepinfra:zai-org/GLM-5.3: 1,000,000 tokens (_AGENTIC)
     • deepinfra:moonshotai/Kimi-K3: 1,000,000 tokens (_AGENTIC_VISION)
     • deepinfra:moonshotai/Kimi-K2.7-Code: 256,000 tokens (_AGENTIC)

3. Automated Tests:

   • tests/test_token_usage.py: verified all 8 DeepInfra models export exact context windows in model_context_windows().
   • tests/test_provider_router.py: verified client builder default base URL and router prefix stripping for DeepInfra models.
   • tests/test_providers.py: verified capabilities (including vision), lockstep reseller descriptor checks, and key isolation (ensuring OPENAI_API_KEY is not leaked when DEEPINFRA_API_KEY is missing).
   • tests/test_provider_verify.py: verified test probe targets https://api.deepinfra.com/v1/openai/models with Bearer auth.

Verification

• Provider and compaction test suites:
  pytest tests/test_token_usage.py tests/test_provider_router.py tests/test_providers.py tests/test_provider_verify.py tests/test_compaction.py tests/test_engine.py -q -> 144 passed.
• Full regression suite: 1943 passed.
• Zero frontend modifications required; the desktop UI automatically consumes model_context_windows from /v1/settings.

Closes #619