Split compose.rs by what it composes

[tauanbinato]

Stacked on #36: its base is large-file-parts, so the diff here is the move alone. Merge #36 first, then retarget this PR to main (branches aren't deleted on merge here, so GitHub won't do it).

src/units/compose.rs was the largest source file (1,915 lines after #36). A blind labeler, reading it with the rest of the codebase, called it a review-level split. Three reasons: the follow-up selectors that only follow_ups.rs uses, the measured caps, and the redundant-test and comment grouping are each read and changed apart, and the project already splits outcome/ and wording/ the same way. I agree, so this moves it into a module:

File	Lines	What it holds
compose/mod.rs	682	compose, the per-rule tally, undecided units, dimensions, file status, the finding
compose/answers.rs	187	the answers an outcome rests on (first pass, rechecks, traces, settles, locates)
compose/due.rs	352	which units each follow-up stage asks next; the functions follow_ups.rs calls are re-exported, so their paths are unchanged
compose/caps.rs	303	the measured limits on a finding's level
compose/located.rs	208	where a finding points: block, value, constant, group or part
compose/redundant.rs	148	redundant tests grouped in threes or more
compose/comments.rs	99	a unit's comments reported together

It's a pure move. Items and their doc comments are unchanged, apart from pub(super) visibility, crate::units:: paths where the code said super::, and Redundant's fields becoming pub(super).

Checked:

• Every one of the 86 items is identical to Ask the parts of long files without a finding whether each does a job of its own #36's compose.rs once visibility, those paths and formatting are normalized (compared mechanically).
• On the whole corpus (181 projects, from the cache), with every clone's token-budget calibration restored before each run, this branch's reports are identical to Ask the parts of long files without a finding whether each does a job of its own #36's head.
• Tests, clippy (also 1.98) and cargo +1.90.0 check --locked pass.

Not done, for later if you want them. The same labeler called four more files consider-level splits:

• analysis/clones.rs: the directory predicates that three other modules import, and the walk-frame recognizer.
• units/security.rs: the settle catalog and Subject with its constructors.
• units/test_units.rs: the setup-text scanning layer.
• units/questions/security.rs: the confirm Choices and the Django checks.

It called file_kind.rs, test_map.rs, transport.rs, docs/load.rs and analysis/comments.rs fine as they are.