fix(model_switch): filter /model picker for providers with no PROVIDER_REGISTRY entry

[Ahmett101]

Summary
list_authenticated_providers() (powers /model in the CLI, gateway, and Telegram picker) emits a row for every slug in PROVIDER_TO_MODELS_DEV whose credential env-var is set. Several of those slugs (notably mistral) have no PROVIDER_REGISTRY entry, so resolve_provider() rejects them with Unknown provider 'mistral' once the user selects a model. The picker shows rows that cannot actually be selected. This is #57503.

Fix: a single resolve-gate in list_authenticated_providers() Section 1 — if PROVIDER_REGISTRY.get(hermes_id) is None, skip the slug. The picker now only lists providers that can actually be switched to. This also resolves the duplicate-Mistral dedup symptom: once the broken-from-models.dev row is filtered, the conflict between PROVIDER_TO_MODELS_DEV['mistral'] and a custom_providers Mistral row becomes moot.

Composes with #50289 (the in-flight first-class-providers plugin merge for mistral): when/if that lands, PROVIDER_REGISTRY gains a mistral entry and the new gate becomes a no-op for it. The two PRs are independent — neither blocks nor breaks the other.

Changes

• hermes_cli/model_switch.py: 9-line addition in list_authenticated_providers() Section 1: a if not pconfig: continue gate after the existing pconfig.auth_type != "api_key" skip. Documented inline, references the issue.
• tests/hermes_cli/test_model_switch_filter_unresolved.py (new): 4 regression tests.
  • test_mistral_filtered_when_unregistered_but_api_key_set — pin: MISTRAL_API_KEY set ⇒ mistral MUST NOT appear in picker.
  • test_resolveable_provider_still_appears — regression guard: deepseek (PROVIDER_REGISTRY-backed) with key set MUST still appear.
  • test_resolve_gate_skips_models_dev_only_provider_without_creds — no-key case stays filtered.
  • test_picker_skips_pconfig_none_does_not_break_other_section1_providers — xai (REGISTRY-backed) with key set still appears.

How to Test

# new regression suite
pytest tests/hermes_cli/test_model_switch_filter_unresolved.py -xvs
# 4/4 passed

# Verify the test catches the bug: revert the fix and re-run
git stash push -m test hermes_cli/model_switch.py
pytest tests/hermes_cli/test_model_switch_filter_unresolved.py::test_mistral_filtered_when_unregistered_but_api_key_set -xvs
# FAILS with "mistral leaked into /model picker: ['mistral']" -- this is what the
# reporter in #57503 saw on the live picker
git stash pop

# Re-run broader suite to confirm no regressions
pytest tests/hermes_cli/test_model_switch_filter_unresolved.py \
       tests/hermes_cli/test_custom_provider_model_switch.py \
       tests/hermes_cli/test_model_switch_custom_providers.py \
       tests/hermes_cli/test_anthropic_picker_curated.py -q
# 55/55 passed

Checklist

• Tests pass -- 4 new + 51 existing model-switch picker tests, 55/55 green
• Follows Conventional Commits (fix(scope):)
• Changes scoped to a single fix, single-file prod change (+1 test file)
• Cross-platform impact: None (os.environ read with monkeypatch, no platform-specific code)
• profile-safe paths: N/A, no paths touched
• .env not used for non-credential settings: N/A, no settings added
• No new dependencies
• Composes cleanly with in-flight feat: add Mistral AI as first-class LLM provider #50289 (different approach, no overlap)

Risk & Impact
Low. Single-file 9-line addition in the picker code path; behaviour change is strictly "drop a row from a list when that row would 100% fail on selection". Users who currently hit the bug get a fixed picker immediately; users with working environments see the picker unchanged (every other provider in PROVIDER_TO_MODELS_DEV is also in PROVIDER_REGISTRY, so they keep appearing).

Type: Bug fix
Closes #57503

Related findings (out of scope here, flagged for maintainers):

• The deeper fix is the provider-plugin promotion of mistral that feat: add Mistral AI as first-class LLM provider #50289 is already driving -- when that lands, the resolve-gate becomes a no-op for mistral and just keeps defending against the next unregistered models.dev provider that anyone adds to PROVIDER_TO_MODELS_DEV. The gate is intentionally defensive: same code path would have caught a follow-up like "provider X has Y key set but isn't in registry" without a second PR.
• A future audit pass over PROVIDER_TO_MODELS_DEV against PROVIDER_REGISTRY (one-shot script, no runtime cost) would let us surface any other slugs that historically slipped through, instead of relying on users to file bugs one-by-one. Out of scope for this PR.

[teknium1]

Closing in favor of the earlier, narrower duplicate #57711 for issue #57503. Both add the same PROVIDER_REGISTRY gate in list_authenticated_providers; #57711 is the minimal production change. Thanks for the reproduction and regression coverage.

[teknium1]

Reopening as the retained implementation for #57503. #57711 is the later duplicate and is now closed; the unselectable models.dev-only provider bug remains reproducible on main, so one focused PR should stay open for salvage.

[teknium1]

Superseded by the improved consolidated salvage PRs prepared in this review batch; contributor credit is preserved in the replacement PR body/commits.