perf!: cache compact DAG analysis

[eitsupi]

Summary

• replace the graph-only manifest cache with a compact model-level analysis containing the DAG, compatibility diagnostics, project name, and referenced file paths
• validate the single cache slot with manifest content bytes under the dlin package-version compatibility boundary
• skip typed Manifest deserialization on warm graph, list, impact, and summary commands while keeping column lineage, MCP, and check-manifest on their direct typed-manifest paths
• restore SQL content from the same bytes used to select the cached DAG, preventing mixed-artifact output when the manifest changes during an invocation
• fail open for corrupt or incompatible cache files and replace cache JSON through a temporary sibling file

Breaking change

The public ManifestGraphCache API is replaced by ManifestAnalysisCache and the non-exhaustive ManifestAnalysis value. dlin is pre-1.0 and there are no known external users of the old raw cache API.

Benchmark

Medium cache workload, 512 models, release binary, 10 runs / 2 warmups. Persistent state is controlled; OS page cache is not flushed.

Manifest summary	Baseline median	Compact cache median
persistent warm	11.599 ms	5.723 ms

The final warm median is about 51% lower than the same-main baseline. The cache file grows from 175,969 bytes to 188,679 bytes (about 7.2%) without persisting the raw typed Manifest. SQL, manifest, and column cold/warm/no-cache semantic probes all pass, and warm cache files remain unchanged.

Validation

• cargo fmt --all -- --check
• cargo clippy --workspace --all-targets --all-features -- -D warnings
• cargo test --workspace --all-targets --all-features
• cache benchmark semantic and invalidation suite on the medium workload