feat(cow): source the registry address from module [config]

[mfw78]

What

Sources the ComposableCoW registry address from the twap-monitor manifest [config] and deletes the compiled cowprotocol::COMPOSABLE_COW constant from the module, with no fallback of any kind.

modules/twap-monitor/src/keeper.rs gains KeeperConfig { registry: Address } with parse mirroring cow-venue's AdapterConfig::parse (unknown keys ignored, malformed value or missing key is a hard Fault::InvalidInput), the same RwLock store/accessor shapes, and a registry field threaded through TwapSource into poll_one's eth_call_params. modules/twap-monitor/src/lib.rs parses and stores in init (a parse failure is the init error); on_block before init is a typed Fault::Unavailable refusal. modules/twap-monitor/module.toml gains [config] with registry = "0xfdaFc9d1902f4e0b84f65F49f244b32b31013b74", the current Sepolia registry, so behaviour is unchanged.

Why

Closes #651: the converged ccow-monitor knows exactly one contract, the registry, and hard-coding the constant coupled the module to a single network and put a cow-rs rev on the critical path. This is the first car of the convergence train and changes the mechanism only; the address cutover to the mainnet fork is #652's atomic swap. A new parity test pins [config].registry to the chain-log subscription address pins and the test fixture constant so the #652 swap must move all three together.

Testing

Gates (all via nix develop -c, from the worktree, package-scoped per the train's light-CI rule):

• cargo fmt --check -p twap-monitor: clean.
• cargo clippy -p twap-monitor --all-targets: clean.
• cargo nextest run -p twap-monitor: 34 passed (25 before, 9 added: config parse valid/missing/malformed/unknown-keys, init-level wiring for the same three cases, a typed-refusal test for on_block without config, a manifest parity test, and a poll-path test that stores a distinct registry and asserts the eth_call to carries it, failing if the constant is reintroduced).
• cargo test -p twap-monitor --lib: 34 passed (the refusal test serialises config access behind a test guard, so the suite is runner-independent).

AI Assistance: Claude Fable used for the implementation via a structured workflow; Claude Opus used for the red-team review (9 findings, 7 fixed, 1 duplicate-merged, 1 rejected as out-of-scope checksum policy).