feat(cli): widen list_models to expose model family (#128, Slice B-i)

[jasonssdev]

Summary

Slice B-i of #128 — the foundation for the interactive model picker. Widens OllamaClient.list_models() so a follow-up slice can filter embedding models out of the chat-model list.

• list_models() now returns list[InstalledModel] (tag + family), parsed from the /api/tags details.family field, instead of a bare list[str].
• New is_embedding_model() classifies the families {bert, nomic-bert} (case-insensitive) as embeddings. A missing or unknown family is treated as non-embedding — ambiguity never hides a legitimate chat model.
• Defensive parsing: a missing / null / non-dict details, or a missing / non-str family, yields family=None without dropping the entry or raising; malformed top-level shapes still surface as OllamaError.
• Every list_models() consumer adapted behavior-preservingly (doctor checks 3/4/5, the init post-write preflight, the model spike). model_tag_matches keeps its list[str] signature; call sites pass [m.tag for m in installed].

Scope

src/openkos/llm/ollama.py (contract), src/openkos/cli/main.py (call-site adaptation), evals/model_spike/run_spike.py (one caller), plus test updates. ~215 lines.

Verification

• uv run pytest: 2099 passed, exit 0
• uv run ruff check . && uv run ruff format --check .: clean
• uv run mypy .: clean
• SDD verify (B-i scope): PASS — 10/10 scenarios
• Adversarial reliability review: 0 findings (parsing robust against hostile /api/tags shapes; all consumers adapted; behavior-preserving)

Strict TDD throughout (RED confirmed before each GREEN).

Issue

Advances #128 (Slice B-i). The picker itself lands in Slice B-ii (a follow-up PR), which will close #128.