chore(config)!: migrate config file format from YAML to TOML

[chaliy]

What

Breaking change: migrate the config file format from YAML to TOML and drop serde_yaml.

• Replace serde_yaml = "0.9" with toml = "1.1" in Cargo.toml.
• Rename Config::from_yaml → Config::from_toml; parser switches to toml::from_str.
• Config::from_file and the --config CLI flag are unchanged — only the file contents differ.
• Convert the two bundled sample configs (benchmarks/config/benchmark.yaml, chaos.yaml) to .toml.
• Update inline test fixtures, the README config block, the architecture spec, the load-test skill, the CLI help text, and benchmarks/run-benchmark.sh.
• Add a CHANGELOG entry with a migration guide.

Why

serde_yaml is published as 0.9.34+deprecated and the upstream dtolnay/serde-yaml repo is archived. Every successor fork is stale (latest releases as of today, 2026-05-17):

Crate	Last release	Status
serde_yaml (current)	0.9.34+deprecated	Archived
serde_yml	2024-08-25	21 months stale; contentious history
serde-yaml-ng	2024-05-26	24 months stale
serde_norway	2024-12-21	17 months stale
toml	2026-04-01	Active — Cargo team, 4 releases in last 30 days

LLMSim's config is a flat, typed, commented dict — exactly TOML's sweet spot. We use zero YAML-specific features (anchors, merges, multi-doc, …), so the migration is mechanical. TOML is also the idiomatic Rust ecosystem config format.

How

• Source change is a one-function rewrite (Config::from_toml); everything else is fixture/doc conversion.
• All four inline test strings were translated YAML→TOML one-to-one.
• Smoke-tested the built binary with --config benchmarks/config/benchmark.toml --port 8888: /health, /openai/v1/chat/completions, /openai/v1/models all respond correctly with the TOML-configured latency=fast, target_tokens=100.

Migration for existing users

- # config.yaml
- server:
-   port: 8080
- response:
-   generator: "lorem"
-   target_tokens: 100
- models:
-   available:
-     - "gpt-5"
-     - "gpt-4o"
+ # config.toml
+ [server]
+ port = 8080
+
+ [response]
+ generator = "lorem"
+ target_tokens = 100
+
+ [models]
+ available = ["gpt-5", "gpt-4o"]

Rules: section: → [section], key: value → key = value, quote strings, YAML lists → TOML arrays. See benchmarks/config/*.toml for full working examples.

Risk

• Medium, surfaced clearly. This is a breaking change for any user with an existing config.yaml. Mitigations:
  • Project is 0.2.x (pre-1.0), so breaking changes are within semver allowance.
  • CHANGELOG includes a side-by-side migration snippet.
  • The --config flag name is unchanged — only the file format moves.
  • Conversion is mechanical and small (typical user configs are ~10-20 lines).

Checklist

• cargo fmt --check
• cargo clippy --all-targets -- -D warnings (clean on rustc 1.95)
• cargo test (166 passed)
• Smoke test with --config benchmarks/config/benchmark.toml
• No serde_yaml references remain outside CHANGELOG explanation
• README, architecture spec, CLI help, and load-test skill updated