fix: correct DeepSeek V4 context windows

[amartejak]

Fixes #389.

Summary

• correct the native DeepSeek V4 Flash and V4 Pro context windows from 128K to the official 1M limit
• record deployment-specific limits for the curated reseller routes: 512K on Together serverless and 1,048,576 on Fireworks and OpenRouter
• add a regression test that pins all five curated DeepSeek V4 context-window values

Using provider-specific limits matters here: DeepSeek exposes the full 1M context natively, while Together's serverless deployment currently exposes 512K. Treating every route as the model-level maximum would make the context meter and compaction trigger overestimate what that endpoint accepts.

Sources checked 2026-08-10:

• DeepSeek: https://api-docs.deepseek.com/quick_start/pricing/
• Together: https://docs.together.ai/docs/serverless/models
• Fireworks: https://fireworks.ai/models/fireworks/deepseek-v4-pro
• OpenRouter: https://openrouter.ai/api/v1/models

Testing

pytest tests/test_token_usage.py tests/test_providers.py tests/test_compaction_engine.py -q
48 passed in 2.09s

No screenshot is applicable: this changes model metadata consumed by the existing context meter and compaction logic, with no UI layout change.