fix(model): warn on stale context_length after /model --global

[LeonSGP43]

Summary

• warn when /model --global leaves an existing model.context_length override out of sync with the newly selected model
• preserve the explicit override instead of auto-rewriting it, while pointing users to /context <n> when they do want to update it
• surface the same warning in both the terminal CLI and gateway /model command, with focused regression coverage

Fixes #10157

Testing

• python3 -m pytest -o addopts="-m 'not integration'" tests/hermes_cli/test_model_switch_context_length_warning.py -q
• python3 -m pytest -o addopts="-m 'not integration'" tests/cli/test_model_switch_context_warning.py -q
• python3 -m pytest -o addopts="-m 'not integration'" tests/gateway/test_model_command_custom_providers.py -q
• python3 -m pytest -o addopts="-m 'not integration'" tests/hermes_cli/test_model_switch_custom_providers.py -q
• python3 -m pytest -o addopts="-m 'not integration'" tests/gateway/test_model_switch_persistence.py -q

[teknium1]

Thanks for preserving the explicit override rather than silently rewriting it. The stale-override premise still exists on current main: cli.py:7946-7949 persists only the model/provider, and the gateway persists the same fields at gateway/slash_commands.py:1916-1923.

Problems

• hermes_cli/model_switch.py:133 uses ModelInfo.context_window before provider-aware resolution. Current resolve_display_context_length() explicitly treats get_model_context_length() as authoritative for provider caps (hermes_cli/model_switch.py:689-715), so this can recommend the vendor-wide context instead of the effective one. The helper also cannot receive custom_providers.
• hermes_cli/model_switch.py:155 recommends /context <n>, but current main has no /context command definition or handler.
• The gateway handler has moved to gateway/slash_commands.py:1399; the current TUI persistence path at tui_gateway/server.py:2795-2812 also needs equivalent handling if this is to cover Hermes' /model surfaces.

Suggested changes

• Rework the helper around the provider-aware resolver with config_context_length=None, and pass compatible custom-provider data.
• Point users to an existing configuration workflow, or implement /context as a separate command.
• Cover provider-capped and custom-provider context values in regression tests.

Automated hermes-sweeper review.

[teknium1]

ModelInfo.context_window can be the vendor-wide catalog value rather than this provider's effective cap. Resolve through the same provider-aware path as current resolve_display_context_length() (and include custom-provider overrides) before composing a recommended replacement.

[teknium1]

/context is not a registered Hermes slash command on current main, so this remediation cannot be followed. Please point to an existing supported config workflow or add the command separately.