Add deterministic agent inference test harness

[Stephen Belanger (Qard)]

This is stacked on #1 and deliberately keeps the agent integration test
harness separate from the daemon migration until it has proven stable. It
should be reviewed and merged after the daemon PR.

The test support is split into reusable layers. server owns the generic Axum
router lifecycle. inference implements programmable OpenAI Responses and
Anthropic Messages routers with captured provider requests. ingest implements
the mock Braintrust API and ordered span/log shape scenarios. agent_process
owns the daemon world and selects inference and ingest backends independently.
Finally, agents contains reusable Codex and Claude Code adapters that hide
installation, isolated configuration, standard CLI flags, mock endpoint
routing, and process output. Run builders still accept extra arguments and
environment variables so future scenarios can cover attachments and other
agent-specific inputs without duplicating setup.

Each agent has one test body for live and mocked modes. Ordinary assertions
always validate stable process behavior and trace delivery. When rows are
locally captured, ordinary assertions also validate stable origin metadata.
IngestScenario is separate and is constructed only when both inference and
ingest are mocked; it validates the additional deterministic request,
tool-result, output, injected-error, and ordered trace-shape expectations.

Inference and ingest modes are orthogonal. BT_AGENT_INFERENCE_MODE selects
mock or live model inference, while BT_AGENT_INGEST_MODE selects captured
local ingest or the normal Braintrust backend. This permits deterministic mock
inference with live Braintrust reporting and no model cost. Mock ingest makes
rows locally available; live ingest waits for emitted spans and fails if the
daemon records a sink error.

CI installs @openai/codex@latest and
@anthropic-ai/claude-code@latest without version pins. The real-agent
scenarios are part of the core daemon matrix on Linux, macOS, and Windows.
This also fixes native Windows Codex compatibility by selecting the
shell_command tool it advertises and by ensuring the .cmd hook shims return
from a bt.cmd capability preflight before forwarding hook input.

Validation:

• cargo fmt --manifest-path bt-daemon/Cargo.toml -- --check
• cargo test --manifest-path bt-daemon/Cargo.toml --all-features
• cargo test --manifest-path bt-daemon/Cargo.toml --all-features --test agent_integration -- --ignored --test-threads=1
• cargo clippy --manifest-path bt-daemon/Cargo.toml --all-targets --all-features -- -D warnings
• make test
• CI runs the daemon and real-agent suite on Ubuntu, macOS, and Windows