feat(cli): add interactive Ollama model picker to init (#128, Slice B-ii)

[jasonssdev]

Summary

Slice B-ii of #128 — the interactive model picker, building on the widened list_models() from #172. Closes #128.

Replaces the free-text model prompt in openkos init with a numbered picker over the chat models actually installed on the local Ollama server, so a first-time user selects from a known set instead of typing an exact tag from memory (the footgun that let a user answer yes in #128).

• _pick_chat_model() probes Ollama in Phase A (before any workspace write), lists installed chat models (embedding models excluded via is_embedding_model), marks the recommended default and selects it on Enter.
• --model <tag> and non-TTY runs bypass the picker unchanged.
• Graceful degradation: an unreachable server or zero chat models falls back to the typed prompt / packaged default; init never hard-fails (workspace still created, exit 0).
• Robust input handling (from adversarial resilience review):
  • the numeric-choice guard requires an ASCII digit — a Unicode digit like ² passes str.isdigit() but is not int()-parseable and previously escaped the bounded reprompt into a hard-fail;
  • the candidate list drops any server tag that would fail validate_model, so the picker never offers an option that hard-fails on selection;
  • the reprompt loop is bounded and falls back to the default after _MAX_PICKER_ATTEMPTS.

Verification

• uv run pytest: 2109 passed, exit 0
• uv run ruff check . && uv run ruff format --check .: clean
• uv run mypy .: clean
• SDD verify (full change): PASS — 8 requirements / 27 scenarios
• Adversarial resilience review: 1 WARNING + 1 SUGGESTION, both fixed in this PR (Unicode-digit hard-fail; picker offering an unvalidatable tag), each with a dedicated regression test.

Strict TDD throughout (RED confirmed before each GREEN, including the two review fixes).

Issue

Closes #128 (Slice B-ii, final slice). Slice A (config hardening, #169) and Slice B-i (list_models widening, #172) already merged.