feat(foundation): bootstrap 0.1.x primitives

[cchinchilla-dev]

What

Bootstrap of the 0.1.x foundation: brings the skeleton from 0.1.0 (contract +
CLI stub only) up to a shape the 0.2.0 MVP can build on top of.

• LLMBackend ABC + Pydantic primitives + parametrised contract suite.
• DirectBackend for OpenAI, Anthropic, Google Gemini via httpx only.
• Pinned pricing.yaml (USD per 1K tokens) + AGENTANVIL_PRICING_FILE override.
• Runner ABC + SubprocessRunner with anyio.fail_after timeout; JSON stdin/stdout protocol.
• Core trace models: Trace, Step, RunRecord, Report.
• Record/replay envelope schema v1 with bit-for-bit determinism CI test.
• AgentAnvilError exception hierarchy.
• py.typed marker, coverage threshold 85%, slow marker excluded by default.
• New CI jobs: version-linearity, determinism.

Why

The 0.1.0 skeleton ships AgentContract and a CLI stub but nothing else.
Every later module needs three primitives in place first:

1. LLMBackend abstraction so backends are swappable.
2. Runner abstraction so the agent under test runs in an isolated process.
3. Bit-for-bit replay envelope so reproducibility holds across machines.

This PR introduces those primitives and pins the contract every later module honours.

Testing

• uv run pytest passes (79 default + 2 slow = 81 tests)
• uv run ruff check src/ tests/ clean
• uv run mypy src/ clean
• python scripts/check_version_linearity.py exits 0
• Portability invariant: DirectBackend imports without AgentLoom
• End-to-end smoke: RecordingBackend -> MockBackend byte-for-byte equal

Notes

• AgentLoomBackend deferred to 0.2.0 (depends on AgentLoom 0.5.0).
• DockerRunner deferred to 0.2.0; K8sRunner to 0.4.0.
• Pricing format mirrors agentloom/providers/pricing.yaml.

[cchinchilla-dev]

Temporary: agentloom constraint lowered from agentloom[contracts,observability]>=0.5.0 to agentloom>=0.4.0 so uv sync works in CI. Revert to the original constraint once AgentLoom 0.5.0 publishes to PyPI.