fix(custom provider): enable prompt caching and accurate context window for custom endpoints serving Claude

[Xiaochengzi2048]

Problem

Two related issues affect users running Claude through a custom OpenAI-compatible endpoint (LiteLLM, Bedrock proxies, local gateways):

1. Prompt caching disabled

anthropic_prompt_cache_policy() returns (False, False) for provider=custom, even when the model name clearly identifies it as Claude. Every request re-bills the full prompt with no cache hits.

2. Context window reported as 200k

lookup_models_dev_context() returns None for provider=custom (not in PROVIDER_TO_MODELS_DEV), so hermes falls back to the 200k default. Claude Opus 4 users see 200k instead of the real 1M window.

Both issues share the same root cause: hermes does not recognise custom as a provider with known model metadata.

Fix

agent/agent_runtime_helpers.py

Add a branch for provider=custom + is_claude that enables caching with envelope layout (same as the existing openrouter branch):

agent/models_dev.py

When provider=custom has no mapping, fall back to the canonical provider based on model name:

This gives Claude Opus 4 its real 1M context window and Sonnet its real 200k — instead of both hitting the 200k fallback.

Related

Closes #50939 (skill delete PermissionError is a separate issue, already in #50940).