feat(init): graceful Ollama first-run in coco init

[gfargo]

First slice of the 0.73 Ollama onboarding polish (freeze-allowed). Since Ollama is the default provider in the picker, this is most new users' first impression of the "local AI commits" path the launch leads with.

Problem

The Ollama branch of coco init shelled out:

ollama list | awk '{print $1}' | awk '{if(NR>1)print}'

• Non-portable — no awk on Windows.
• Crashes if Ollama isn't installed (execPromise rejects with a raw "command not found").
• On no models, prints a vague message and hard-exits, discarding every prior init answer (scope, commitlint).
• No "daemon not running" guidance.

Change

New lib/langchain/utils/ollamaStatus.ts:

• getOllamaStatus(endpoint?) — probes the HTTP API (/api/tags) for reachability and pulled models in one cross-platform call; a PATH probe only distinguishes "not installed" from "not running" for the copy.
• pullOllamaModel(model) — ollama pull with live progress.
• OllamaNotReadyError — typed signal.

coco init now guides each first-run state instead of crashing:

• not installed → install link (ollama.com/download / brew install ollama)
• installed but not running → ollama serve
• up but no models → recommends + optionally pulls llama3.1:8b (~4.7 GB, the cost-tier commit default), prompted — never automatic

When Ollama can't be used and the user doesn't fix it inline, the handler catches OllamaNotReadyError and re-offers the provider picker — prior answers are preserved (no more session loss).

Tests

• ollamaStatus.test.ts — reachable / not-running / not-installed / non-OK / endpoint normalization (7 cases).
• init.test.ts — new test: not-ready → re-prompt → completes with another provider.

Validation

eslint (0 errors), tsc --noEmit clean, jest init + ollamaStatus suites green (7/7 each). Full rollup build deferred to CI (nested-worktree PATH quirk locally).

Follow-ups (rest of the 0.73 Ollama polish)

• coco doctor: Ollama daemon liveness + configured-model-pulled check
• Actionable runtime error on commit when daemon down / model missing
• Fix handleMissingApiKey Ollama copy (OLLAMA_API_KEY is misleading for local)