feat(ui): one type scale, and one estimate of how wide a character is

[youhide]

T2. Eight sizes across 249 call sites, and six file-local constants duplicating each other — ITEM_SIZE = 13.0 written out three separate times in three widgets, CODE_SIZE = 12.0 three more. All of them are now steps on a named scale.

Six steps, named for the job

MICRO LABEL CODE BODY LEAD DISPLAY. LABEL carries structure — section headings and secondary text, the commonest size in the interface — and BODY carries content.

The guard is split by scale, not by file

The two scales migrate at different speeds. A widget can be entirely on the type scale while its paddings are still hand-picked, and one list for both would force converting everything about a file at once or nothing at all.

So ON_THE_LAYOUT_SCALE and ON_THE_TYPE_SCALE are separate lists over separate property sets, and a test pins that one guard does not see the other's numbers.

Two sizes deliberately left out — and why that is a finding

14.0 and 16.0 are each a third spelling of "the title of something", a pixel away from LEAD.

The settings panel shows why that is a hierarchy question rather than a rename:

its header is 14.0 while the option labels beneath it are 15.0 — the heading is smaller than what it heads.

Deciding what those sites should say is a visual decision, not a substitution, so it is not in this change. The six files holding a 14 or a 16 stay off the type list until it is made. That is also why this PR has no visual change at all: every size that moved kept its value, and the ones that could not are excluded rather than approximated.

One estimate of character width

The canvas has no cheap text measurement, so format::truncate and the graph's badge_width both estimate. They estimated differently — 0.62 against 0.58, for the same font.

The smaller one is the graph's, which is the direction that hurts: a badge too narrow for a label that truncation had already called short enough. And badge_width's own doc comment warns that "two copies of this arithmetic would drift" — the ratio was quietly the second copy.

Pinned by a compile-time assertion rather than a test:

const _: () = assert!(CHAR_WIDTH >= crate::format::CHAR_WIDTH_RATIO);

Clippy is right that a runtime assertion over two constants says nothing, and a wrong value here should fail the build, not a run. I wrote it as a test first, clippy caught it, and the const assertion is the better answer.

Sabotage — three, all confirmed

Sabotage	What broke
a raw .size(11.0) returns to sidebar.rs	migrated_files_carry_no_raw_type_sizes
a raw padding returns to tabs.rs	migrated_files_carry_no_raw_layout_numbers, proving the split lists still work
the graph's ratio goes back to 0.58	compile error — evaluation panicked: assertion failed

Docs

UI_SPEC.md's Layout scale section gains a Type subsection: the six steps, the measurement that motivated them, the 14/16 hierarchy problem stated plainly, and the character-width reconciliation.

Gate

cargo fmt --all -- --check, cargo clippy --workspace --all-targets --all-features --locked -- -D warnings, cargo test --workspace --all-features --locked — 895 tests, all green.