fix(runtime): break divert target reference cycles

[bladecoder]

Summary

• replace the cached strong Rc<Container> held by Divert with a Weak<Container> plus the target index
• preserve lazy target resolution while preventing self- or ancestor-targeting diverts from retaining the story graph
• add a focused unit test that constructs the reference cycle directly and verifies that the container is released

Context

This supersedes #23 and retains Brent Anderson’s original diagnosis and core Weak-cache approach. The implementation has been ported to the current runtime/ workspace layout and updated for the repository’s current formatting and Clippy requirements.

The original allocator-based regression test ran 1,000 full stories, took roughly 9–13 seconds locally, and retained about 667 MB when demonstrating the failure. The replacement test checks the ownership invariant directly, avoids a global unsafe allocator, and completes immediately.

Validation

• cargo fmt --all
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace
• cargo test -p bladeink --features stream-json-parser