fix(ollama-cloud): preserve :cloud tags for local-daemon proxy in model catalog

[allenramirez-arch]

Problem

When Ollama Cloud is routed through the local Ollama daemon (base_url host localhost / 127.0.0.1 / ::1 / *.local), cloud models must keep their Ollama tag suffixes:

• kimi-k2.6 → kimi-k2.6:cloud
• qwen3-coder:480b → qwen3-coder:480b-cloud

fetch_ollama_cloud_models() merged the live list with the models.dev catalog and ran every models.dev ID through _strip_ollama_cloud_suffix(), which unconditionally removed :cloud/-cloud. The merged/cached list (and therefore the hermes model picker) then offered bare IDs that the local proxy cannot serve.

Fix

Replace the unconditional strip with intent-separated helpers:

• _ollama_cloud_suffix_key() — suffix-agnostic value used only as a dedup equivalence key (never saved as a model ID).
• _ollama_cloud_base_is_local_proxy() — detects local-daemon routing.
• _ollama_cloud_model_for_endpoint() — re-tags catalog IDs for the local proxy (name → name:cloud, name:tag → name:tag-cloud) while preserving hosted ollama.com/v1 IDs as-is.

The merge keeps the real, tagged IDs and uses the stripped value only for equivalence, so bare/suffixed duplicates still collapse to one entry.

Tests

Extends tests/hermes_cli/test_ollama_cloud_provider.py (45 passing locally).

[alt-glitch]

This was generated by AI during triage.

Duplicate of #52613 — both patch fetch_ollama_cloud_models() in hermes_cli/models.py to use _strip_ollama_cloud_suffix only as a dedup equivalence key and preserve the real tagged model ID, fixing the same over-stripping bug from #52599. #52613 is the earliest open version and is canonical; this PR adds extra local-proxy re-tag helpers but the core mechanism is identical.

[teknium1]

Thanks for tracing the suffix loss. The premise is still present on current main: hermes_cli/models.py:4067-4070 strips each models.dev ID before adding it to the picker/cache.

Problems

• The proposed endpoint-shaped IDs need endpoint-aware caching. Current fetch_ollama_cloud_models() returns the shared disk cache before resolving base_url (hermes_cli/models.py:4032-4042), while _save_ollama_cloud_cache() stores only models and a timestamp (hermes_cli/models.py:4005-4011). A local-proxy refresh could therefore make a later hosted request reuse tagged IDs, or vice versa.

Suggested changes

• Key the cache by endpoint shape/base URL, or transform cached canonical IDs after resolving the selected endpoint.
• Add a cache-hit regression covering a local-proxy call followed by a hosted call (and the reverse), not only force_refresh=True.

Automated hermes-sweeper review.