fix: honor custom provider context in web model info

[qsunyi]

Summary

• make /api/model/info honor per-model context_length from custom providers
• add a regression test covering a custom provider like roy/gpt-5.4

Test Plan

• bash scripts/run_tests.sh tests/hermes_cli/test_web_server.py -k custom_provider_context_length
• bash scripts/run_tests.sh tests/hermes_cli/test_web_server.py

[teknium1]

Thanks for covering a real dashboard/runtime mismatch. Current main still has the gap: get_model_info() reads only model.context_length at hermes_cli/web_server.py:5140 and calls metadata resolution without custom providers at hermes_cli/web_server.py:5153-5158.

Problems

• The added resolver duplicates the canonical get_custom_provider_context_length() helper at hermes_cli/config.py:5092. Its break paths differ from the shared helper's malformed-entry handling at hermes_cli/config.py:5133-5153, so the two paths can disagree.
• The new elif provider guard is narrower than runtime resolution: agent/agent_init.py:1668-1678 resolves matching custom-provider overrides from model and base URL without that guard.
• The regression test covers providers only. The shared helper supports legacy custom_providers as well (hermes_cli/config.py:5118-5125).

Suggested changes

• Reuse get_custom_provider_context_length(model=model_name, base_url=base_url, config=cfg) directly and report its result as the config-derived override.
• Add the same endpoint contract for both compatible configuration forms.

The linked #10690 discussion also recommended consolidating this bug class around the shared resolver. This is an automated hermes-sweeper review.