## What

Dependency incompatibilities now store only the dependent and dependency
package IDs in `Kind::FromDependencyOf`. The corresponding version sets
are moved directly into the incompatibility terms instead of being
cloned into both representations, and the unchanged four-field
`External::FromDependencyOf` report is reconstructed from those terms on
the cold error-reporting path.

The latest `main` branch also indexes merge candidates by
dependency-range hash. To preserve that behavior without restoring a
clone or adding another field to every incompatibility, `as_dependency`
borrows the optional dependency range from the terms and the merge cache
hashes that optional reference. Empty dependencies remain represented by
an omitted negative term; self-dependencies, duplicate dependencies,
merged dependents, iteration order, and public `Incompatibility::iter`
behavior are preserved.

This changes the public `Kind::FromDependencyOf` variant from four
fields to two, so it is intended for a 0.5.0-or-later release. [The
dependent uv draft](astral-sh/uv#20048) updates
the four consumer matches and temporarily pins this commit.

## Why

`Incompatibility::from_dependency` is on the dependency-construction hot
path. Previously, it cloned both version sets into `Kind` even though
the same sets were already owned by `package_terms`. With clone-counting
coverage, the constructor now performs zero `VersionSet` clones for both
empty and non-empty dependencies.

The clone-removal mechanism was measured against the independent
[reserve-only control](#68)
over five CodSpeed CPU-simulation samples:

| Benchmark | Reserve-only control | Clone removal | Change |
| --- | ---: | ---: | ---: |
| `sudoku-easy` | 4.70 ms | 3.55 ms | 24.47% faster |
| `sudoku-hard` | 4.86 ms | 4.17 ms | 14.20% faster |

Against the original baseline, the cumulative candidate was 26.35%
faster on `sudoku-easy` and 18.87% faster on `sudoku-hard`.
`backtracking_ranges` moved from 1.89 s to 1.90 s, a disclosed 0.53%
regression; the other measured benchmarks were neutral to modestly
faster. [Representative CodSpeed
report](https://app.codspeed.io/astral-sh/pubgrub/runs/6a41c50ad94dafe8c8334e3a).

The full workspace and focused semantic tests pass, including zero-clone
construction, empty and self-dependencies, duplicate term ordering,
derivation-tree reconstruction, merged ranges, and package
self-dependency behavior. Strict Clippy, formatting, documentation, and
diff checks also pass. The small latest-main merge-cache adaptation is
integration-only and was not measured separately from the validated
clone-removal mechanism.