Module plan 

Tier 2 step 1 — the cohesion clusterer.

This is not a bin-packer. The objective is to put related items together and unrelated items apart. The number of output files is a consequence of how related the code is, never something to minimise: if a file genuinely contains ten unrelated 50-line items, ten 50-line files is the correct answer, and merging any of them to “save a file” or “fill a bin” makes the split worse. max_loc is only a ceiling, never a packing target: a transitively related group bigger than half the ceiling is partitioned into roughly equal cohesive parts (strongest reference edges bond first, so the cuts fall on the weakest edges), leaning toward more files rather than files that graze the ceiling.

Relation: an item references a sibling (the manifest’s adjacency_hint). Clustering is greedy agglomerative on those reference edges, heaviest edge first, capped at the component’s balanced share; leftover clusters of the same component are then consolidated in source order up to the same share. Items with no edges stay alone. Imports/preamble are pulled out as a shared header (the reassembler copies into each module only the imports it references); an item that is itself >= max_loc is reported oversized.

Three chunk classes never enter clustering:

• visibility-qualified use/extern crate (pub use ...) are re-exports — the file’s API surface — and stay at the root verbatim (root_items);
• content-less mod name; declarations bind files relative to the root’s directory and stay at the root verbatim (root_items);
• body-carrying mod name { ... } items are already module boundaries and are handled separately (mods). Unconditional modules are extracted; conditional modules retain their braced scopes inline.

Structs

Oversized

An item whose own LOC is at or above the ceiling.

Part

One cohesive module: a cluster of related items.

SplitPlan

The result of planning a split.

Functions

plan_split

Cluster exploded’s items into cohesive modules, each < max_loc.