Adopt Dagger for CI pipeline, replacing docker-compose

[coilysiren]

Why

Gauntlet is the best-fit repo in the coilysiren portfolio to adopt Dagger:

1. Already a multi-container pipeline. docker-compose.yml has api, demo, and test services with healthcheck-gated dependencies. This is exactly what Dagger replaces — the compose topology becomes a typed Python function where test depends on api.as_service(), with real caching between runs.
2. uv already in use. Dagger's Python SDK is first-class and plays nicely with uv. pyproject.toml needs no changes.
3. Dagger's LLM/agent primitives map onto what Gauntlet is. Dagger shipped LLM-native pipeline nodes in 2025 — pipeline steps can call Anthropic/OpenAI with tool-use, sandboxed per run. Gauntlet already orchestrates Anthropic + OpenAI as attacker/inspector; running those inside Dagger gives per-run caching, parallelism, and a trace UI for free.
4. Matrix runs become trivial. The (attacker × inspector × target) combinatorial explosion is ugly in compose or pytest. In Dagger it's a map over a list of configs.
5. Local = CI. Same function runs on laptop and in GitHub Actions. No duplicated pipeline definitions, no "works on my machine."

Scope

• Add dagger-io to dev dependencies via uv
• Create ci.py (or gauntlet/ci/) with a dagger.Client-based module
• Port api service → Dagger function returning a Service
• Port demo service → Dagger function that consumes api
• Port test service → Dagger function running pytest -m "not docker"
• Add a top-level test() function composing the three
• Add a matrix function for (attacker_type × inspector_type) combinations
• Wire into GitHub Actions as a single dagger call test step
• Delete docker-compose.yml once parity is reached
• Update README.md / CONTRIBUTING.md with the new dev loop

Non-goals

• Not adopting Dagger in other coilysiren repos yet (backend is a candidate second, but after this beachhead proves out)
• Not replacing invoke/tasks.py workflows outside CI

Context

• Upstream: https://github.com/dagger/dagger
• Python SDK: https://docs.dagger.io/sdk/python
• Motivation is also career/platform-eng signal — Dagger is trending in 2025-2026 alongside Backstage as a "programmable CI" story.