fix(zai): add fetch_models override for GLM-5.2 + base_url fix

[ether-btc]

Problem

GLM-5.2 and several GLM-4.5 variants work via /chat/completions but are absent from the /models catalog endpoint. Without a fetch_models override, the model picker silently drops these models — the user can't select them even though they're functional.

Changes (1 file, +40/-2)

plugins/model-providers/zai/__init__.py:

1. Add ZAIProfile(ProviderProfile) with a fetch_models() override that:

   • Fetches the live catalog via super().fetch_models()
   • Merges it with a curated _EXTRA_MODELS set (8 known-working models)
   • Falls back to the full _EXTRA_MODELS list if the catalog is unreachable

2. Fix base_url: /api/paas/v4 → /api/coding/paas/v4

3. Expand fallback_models: added glm-4.5, glm-4.5-air, glm-4.5-flash

Verification

Tested on a Raspberry Pi 5 running Hermes Agent. GLM-5.2 is the daily driver model — it works reliably via /chat/completions but was invisible in the model picker until this fix was applied.

---

Replaces #46615 (closed — that branch carried 146 unrelated file changes from a stale fork point). This PR is a clean cherry-pick against latest main.

[teknium1]

Thanks for the focused cleanup. The primary picker regression has been addressed on current main by the shared catalog path: 630b43892d7e795f7ebf84b0d9ea8f0428a3692b added live-plus-curated merging, now implemented at hermes_cli/models.py:2485-2517; the Z.AI curated list already includes glm-5.2, glm-4.5, and glm-4.5-flash at hermes_cli/models.py:299-308.

Problems

• The proposed ZAIProfile.fetch_models() does not accept base_url, but current main calls it with base_url=... at hermes_cli/models.py:2485. Applied unchanged, authenticated Z.AI picker discovery would raise TypeError; it also would not forward a resolved custom endpoint to ProviderProfile.fetch_models().
• Changing plugins/model-providers/zai/__init__.py's base_url does not change the runtime Z.AI default: runtime resolution uses hermes_cli/auth.py:243-250 and hermes_cli/auth.py:6462-6463.

Suggested changes

• Re-scope to any verified remaining IDs absent from the shared curated Z.AI catalog (for example, glm-4.5-air / glm-4.6 if confirmed), rather than duplicating the generic merge mechanism.
• If an override remains necessary, preserve and forward base_url and add a focused regression test.

Automated hermes-sweeper review.

[teknium1]

Current main calls provider profiles as _p.fetch_models(api_key=..., base_url=...) in hermes_cli/models.py:2485. This override must accept base_url and forward it to super().fetch_models(); otherwise authenticated Z.AI picker discovery raises TypeError and ignores a resolved custom/coding endpoint.

[teknium1]

This profile field is not the runtime Z.AI endpoint source: resolve_api_key_provider_credentials() derives the default from hermes_cli/auth.py's PROVIDER_REGISTRY and _resolve_zai_base_url(). Changing this line alone will not route normal Z.AI inference to the coding endpoint.