perf(zai): parallelize endpoint detection probes

[light-merlin-dark]

Problem

Z.AI has separate billing for general vs coding plans and global vs China endpoints. On startup, detect_zai_endpoint() probes up to 4 endpoints sequentially with 8s timeout each.

When the first endpoints return non-200 (e.g. rate-limited 429), the probe walks through all of them before finding a working one. This adds 8-9 seconds to every cold startup where the endpoint cache doesn't exist.

Solution

Replace the sequential for loop with concurrent.futures.ThreadPoolExecutor to probe all 4 endpoints in parallel. Results are collected and returned in ZAI_ENDPOINTS priority order so the preference chain (global → cn → coding-global → coding-cn) is preserved.

Metrics

Measured on macOS M4 Max, Python 3.11, Hermes v0.8.0:

Endpoint	Sequential timing	Status
global	0.9s	429 (rate limited)
cn	1.6s	429 (rate limited)
coding-global	4.3s	200 ✓
coding-cn	2.0s	200 ✓

	Before	After
Total probe time	8.8s	~4.5s

The parallel version is bounded by the slowest single endpoint rather than the sum of all endpoints.

Files changed

• hermes_cli/auth.py — detect_zai_endpoint() now uses ThreadPoolExecutor; extracted _probe_single_zai_endpoint() helper.

---

Merlin (@EnchantedRobot on X) & GLM 5.1 via OpenCode

[teknium1]

Thanks for targeting a real cold-start latency path: current main still probes synchronously in hermes_cli/auth.py:639-667.

Problems

• Current main changed ZAI_ENDPOINTS so its third field is probe_models, a list (hermes_cli/auth.py:623-628). The proposed helper treats that field as one scalar model (hermes_cli/auth.py:392) and posts it directly (hermes_cli/auth.py:401), which would send a list and lose the current fallback across candidate models.
• There is no direct probe test. Existing Z.AI tests monkeypatch detect_zai_endpoint() at the credential-resolution boundary (tests/hermes_cli/test_api_key_providers.py:963-1007).

Suggested changes

• Salvage as one worker per endpoint that preserves the current per-endpoint candidate-model loop, then select successful endpoint results in ZAI_ENDPOINTS priority order.
• Add mocked coverage for candidate fallback and deterministic endpoint-priority selection.

Automated hermes-sweeper review.

[teknium1]

Current main's third tuple element is probe_models, a list (hermes_cli/auth.py:623-628), not a scalar. A salvage of this helper would send that list as json["model"] and skip the existing candidate-model fallback. Please make the worker iterate each endpoint's candidate models and return the first successful scalar model.