feat(instrument): Port Ollama LLM provider adapter (M3)

[mmercuri]

Summary

M3 LLM-provider fan-out — self-contained PR that lands the Ollama
adapter together with the minimum base infrastructure required to
host it.

• Source: ateam/stratix/sdk/python/adapters/llm_providers/ollama_adapter.py (~259 LOC)
• Sister M3 PRs (in parallel): OpenAI, Anthropic, Azure OpenAI, AWS Bedrock, Google Vertex, LiteLLM, Cohere, Mistral
• Template: matches the structural shape used by the OpenAI adapter port (providers/openai_adapter.py + flat _base/)
• Local-only: Ollama runs on the operator's own hardware. Default endpoint is http://localhost:11434 (overridable via OLLAMA_HOST)
• Pricing: every Ollama model is recorded as 0.0 USD/token in the canonical pricing manifest with an inline comment explaining why (self-hosted). Optional infra_cost_usd derived from compute duration when cost_per_second is configured by the caller. Token counting is still required — the adapter records prompt_eval_count / eval_count from the daemon response on every call.

What's in the box

Deliverable	Path
Adapter	src/layerlens/instrument/adapters/providers/ollama_adapter.py
Lazy public export	src/layerlens/instrument/adapters/providers/__init__.py (__getattr__)
Pricing entries (zero-cost)	src/layerlens/instrument/adapters/providers/_base/pricing.py
Tests (respx HTTP fixtures)	tests/instrument/adapters/providers/test_ollama_adapter.py
Sample (mocked + live modes)	samples/instrument/ollama/{__init__.py,main.py,README.md}
Docs (incl. ollama serve setup + GPU notes)	docs/adapters/providers/ollama.md
pyproject extra	providers-ollama = ["ollama>=0.2"]
Base infrastructure	_compat/, instrument/_vendored/, instrument/adapters/_base/, instrument/adapters/providers/_base/

Acceptance

• uv run pytest tests/instrument/adapters/providers/test_ollama_adapter.py -x — 13 passed
• uv run pytest tests/instrument/test_lazy_imports.py — 3 passed
• uv run pytest tests/instrument/test_default_install.py — 3 passed
• uv run mypy --strict src/layerlens/instrument/adapters/providers/ollama — clean (6 source files)
• uv run ruff check src/ tests/ — clean
• Lazy-import + default-install guards still pass

Test plan

• Reviewer confirms ollama is NOT loaded by import layerlens (covered by test_lazy_imports.py::test_layerlens_import_does_not_pull_frameworks)
• Reviewer confirms pip install layerlens (no extras) does not change the runtime dep set (test_default_install.py)
• Reviewer optionally runs LAYERLENS_OLLAMA_LIVE=1 python -m samples.instrument.ollama.main against a local ollama serve daemon
• Reviewer confirms api_cost_usd is exactly 0.0 on every emitted cost.record event (self-hosted invariant)
• Reviewer confirms the pricing-manifest zero-cost entries match the canonical Ollama tag list