fix(ollama): preserve cloud model suffixes in picker

[deranaz]

What does this PR do?

Preserves Ollama Cloud model IDs from models.dev in the model picker output while still using suffix-stripped IDs only for deduplication. This prevents cloud-only IDs such as qwen3-coder:480b-cloud from being shown as local-looking IDs like qwen3-coder:480b.

Related Issue

Fixes #52599

Type of Change

• Bug fix
• Tests

Changes Made

• hermes_cli/models.py: keep the original Ollama Cloud model ID when adding merged entries, but dedupe by the normalized suffix-stripped key.
• tests/hermes_cli/test_ollama_cloud_provider.py: update suffix tests to assert cloud IDs are preserved when models.dev is the source, and add coverage for live/API dedupe with suffixed IDs.
• tests/hermes_cli/test_ollama_cloud_provider.py: add explicit UTF-8 encodings in cache test file I/O to satisfy the Windows compatibility checker.

How to Test

1. Run uv run --extra dev pytest tests/hermes_cli/test_ollama_cloud_provider.py tests/hermes_cli/test_setup_ollama_cloud_force_refresh.py -q.
2. Run uv run python scripts/check-windows-footguns.py --diff upstream/main.
3. Run git diff --check.

Checklist

Code

• I read the contributing guide.
• Commit message follows Conventional Commits: fix(ollama): preserve cloud model suffixes in picker.
• I searched existing issues/PRs before opening this PR.
• PR is focused on one logical bug fix.
• Added/updated regression tests for the bug.
• Platform tested: macOS, focused Python test suite.

Documentation & Housekeeping

• Documentation update: N/A, no user-facing docs changed.
• cli-config.yaml.example: N/A, no config keys changed.
• CONTRIBUTING.md / AGENTS.md: N/A, no architecture/workflow changed.
• Cross-platform impact considered; scripts/check-windows-footguns.py --diff upstream/main passes.
• Tool descriptions/schemas: N/A, no tool schema changed.

Screenshots / Logs

46 passed in 11.59s
✓ No Windows footguns found (2 file(s) scanned).

[teknium1]

Thanks for the focused regression coverage. The current merge does still normalize models.dev-only IDs, but this patch reverses a documented compatibility decision that needs endpoint-specific handling.

Problems

• hermes_cli/models.py:3615 changes the saved/returned models.dev ID globally. Current main deliberately stores the normalized ID at hermes_cli/models.py:3767-3772; commit eadf34633e038c595dcf615845b992a45443e380 records that the hosted Ollama Cloud API uses clean IDs and rejected suffixed ones with 400/404 responses.
• The patch has no base_url distinction. The related local-proxy report in #59819 identifies routing through a local Ollama daemon as the case where tags may be required, which is different from the hosted default at hermes_cli/models.py:3742.

Suggested changes

• Preserve the hosted-endpoint normalization, and make any tag preservation/retagging conditional on a verified endpoint mode.
• Cover hosted and local-proxy merge/cache results separately.

Automated hermes-sweeper review.

[teknium1]

This globally changes the cached picker ID, but current main intentionally appends the normalized value because commit eadf346 documents hosted Ollama Cloud rejecting suffixed IDs. Please scope tag preservation to a verified endpoint mode (for example, a local daemon proxy) rather than applying it to every models.dev fallback.