fix: align Fireworks model picker with setup flow

[ft-ioxcs]

Summary

Fixes a Fireworks model-picker discrepancy between the standalone setup flow and the in-session /model picker.

The setup flow (hermes setup model / hermes model) shows Fireworks models from the curated models.dev agentic list, with routers like:

accounts/fireworks/routers/kimi-k2p6-turbo
accounts/fireworks/routers/kimi-k2p7-code-fast
accounts/fireworks/routers/glm-5p1-fast

But the in-session /model picker was going through the generic live provider catalog path first. For Fireworks, that live list could start with older/live-only model IDs such as:

accounts/fireworks/models/cogito-v1-preview-llama-3b

So the two model-selection surfaces appeared to offer different Fireworks model lists.

Change

For generic API-key providers that have no static curated list but are marked as models.dev-preferred, use the models.dev agentic list as the curated base before appending live-only API models.

This keeps Fireworks in the same top/order as the setup flow while still preserving live-only discovery after the curated agentic models.

The change is intentionally narrow:

• Existing providers with non-empty static curated lists keep their previous curated-first behavior.
• Non-models.dev-preferred providers with no curated list still return live models as-is.
• Fireworks now gets models.dev entries first, matching setup-model order.

Test plan

Confirmed the new regression test fails before the fix.

Then ran:

./venv/bin/python -m pytest tests/hermes_cli/test_provider_live_curated_merge.py -q

Result:

10 passed

Ran the model-options parity guard too:

./venv/bin/python -m pytest \
  tests/hermes_cli/test_provider_live_curated_merge.py \
  tests/test_tui_gateway_server.py::test_model_options_does_not_overwrite_curated_models \
  -q

Result:

11 passed

Compiled touched Python files:

./venv/bin/python -m py_compile hermes_cli/models.py tests/hermes_cli/test_provider_live_curated_merge.py

Ran a live local comparison against the current Fireworks catalog path:

setup_count 15
provider_count 28
payload_count 28
first15_match True

The /model payload now starts with the same first 15 Fireworks entries shown by the setup flow, including the Fireworks routers from the screenshot.

[teknium1]

Fireworks AI support landed via PR #62593. This contribution was part of the Fireworks provider cluster credited in the salvage PR; thank you for the implementation and review work.