feat(model): honor discover_models in terminal hermes model named-custom flow (closes #18726)

[kshitijk4poor]

Summary

Fixes #18726 — the terminal hermes model wizard (_model_flow_named_custom) always live-probed a custom provider's /models endpoint and ignored the configured models: list. For plans whose endpoint exposes a large catalog this floods the picker with models the user can't use (the issue's example: Baidu Qianfan Coding Plan returns 100+ models — ERNIE, embeddings, image models — for a 2–3 model plan).

Root cause

Two gaps in the terminal flow (the slash-command /model picker in model_switch.py §3/§4 already handles this):

1. _named_custom_provider_map() built the provider_info dict but dropped both models and discover_models, so the flow never received them.
2. _model_flow_named_custom() called fetch_api_models() unconditionally.

Fix (Option B — explicit opt-out, consistent with §3/§4 and #25810/#40080)

• Wire models + discover_models through _named_custom_provider_map into the flow.
• Honor discover_models: false: use the configured models: list verbatim and skip the live probe. String values ("false"/"no"/"0") normalise to False, matching the slash picker.
• Default stays True — no behaviour change for existing configs. A models: list alone does not implicitly whitelist (that would silently break the documented "configure models: for context-length overrides while still live-fetching" use case).
• If discovery is on but the probe returns empty, fall back to the configured models: list instead of forcing manual entry.

Config (now works)

providers:
  baidu:
    base_url: https://qianfan.baidubce.com/v2/coding
    key_env: BAIDU_CODING_API_KEY
    discover_models: false
    models: [kimi-k2.5, glm-5]

hermes model → Baidu → shows exactly 2 models, no probe.

Why not the implicit-whitelist default

The issue's literal ask was "no flag needed". I went with the explicit discover_models: false opt-out instead because: (1) it's one consistent mental model across every picker path (discover_models: false = "trust my list" everywhere — §3, §4, and now the terminal flow); (2) flipping the default would silently break configs that list models: purely for context-length overrides while relying on live discovery. The Baidu user adds one line and gets exactly the requested result.

Tests

tests/hermes_cli/test_custom_provider_model_switch.py — new TestCustomProviderDiscoverModels (5 tests):

• discover_models: false + configured models → no probe, list used
• selection from the configured list persists
• default (flag unset) still probes even with a models: list (Option B guard)
• probe-empty falls back to the configured list
• string "false" normalises to disable

All pass. Verified end-to-end with a live call that raises if fetch_api_models is touched under discover_models: false — probe is skipped, "Using configured models" shown, exactly the 2 configured models listed.

Note: 3 unrelated tests in test_model_switch_custom_providers.py (groups_same_endpoint, distinct_endpoints_stay_separate, total_models_reflects_grouped_count) fail identically on bare origin/main — they don't mock fetch_api_models and probe a live local Ollama. Not touched by this PR.