fix(cli): preserve static custom provider models + prune stale custom credentials (2-PR salvage)

[teknium1]

Summary

Two config/auth fixes: custom providers declaring models: as dict rows ([{id: ...}]) no longer show zero models in the picker, and stale model_config credential mirrors are pruned when the user switches custom endpoints (stops requests going to the old endpoint).

Salvages #58453 (@lord-dubious) and #58409 (@msh01) onto current main, authorship preserved. The cron per-job max_tokens PR (#58440) originally clustered here was dropped by maintainer decision — per-job token caps are not a user-facing knob we want.

Changes

• hermes_cli/model_switch.py + hermes_cli/config.py: new _declared_model_ids() helper accepting string and dict (id/name) rows, replacing 4 divergent ad-hoc parsers (refs Desktop model picker cannot reliably show or select configured custom OpenAI-compatible providers #58393)
• hermes_cli/model_setup_flows.py: after an explicit custom-endpoint save, prune source=="model_config" entries from inactive custom:* pools; manual entries retained (refs [Bug]: auth.json caches credentials, causing double requests, old and new endpoints #57569)
• Regression tests for both

Validation

	Before	After
custom provider with dict-row models	picker shows (0) models	models listed
switch custom endpoint	old-endpoint cred can still be selected	stale mirror pruned

Targeted suites: inventory, model_switch custom providers, model_setup_flows — all pass.

Closes #58453. Closes #58409.

Infographic