feat(diff): highlight where the walkthrough points, dim what a rule can prove is mechanical

[fiddur]

Attention is the scarce thing in a review, and a diff spends it evenly. Two mechanisms, with the
decider deliberately different for each.

Highlighting is the walkthrough's job. A tour step already records the lines it is about, so the
hunks covering them are tinted. The AI only ever adds attention — it can never take any away.

Dimming is a rule's job, never a model's. Dimming asserts that something needs less attention,
which is the one claim a review tool should have to show its reasoning for. A hunk is dimmed when:

• every line it touches is an import (ES, require, Python, Rust, #include, Ruby, Go/Java package lines)
• its added and removed lines are the same lines with different whitespace
• the file is generated (reuses the existing lockfile/generated detection)

It recedes to 45% and comes back on hover, stays selectable and commentable, and carries the reason
as a title, so a reader can always find out why instead of having to trust it. A dim you cannot
interrogate is the tool lying about the size of the change.

Indentation-sensitive files are never whitespace-dimmed — .py, .yml, .yaml, .md,
Makefile, .hs and friends. There the indentation is the syntax, so a reindent can change which
key a value hangs off or which block a statement belongs to. There is a test asserting a YAML
reindent stays at full attention.

A deliberate conservatism: an added or removed blank line does not count as whitespace-only,
because it changes the file's shape. git diff -w agrees — it needs --ignore-blank-lines for that.

Whitespace hiding now defaults on and is remembered, since it is the formatter's business
(prettier, oxfmt). The toolbar appends · whitespace hidden to the description whenever it is on:
a filtered diff renders fewer lines than GitHub does, and after the +14524-vs-+367 episode that
must never be silent.

Threading: attentionClass and attentionTitle reach every row group of a hunk. The hunks render
inside a <table>, so there is nowhere to hang a wrapper element — the class goes on each <tbody>
the hunk emits, in both the unified and split renderers.

152 UI tests pass.

Two things left for the follow-up: scrollToLine, so the stepper lands on a step's lines rather
than the top of its file — there are no per-line DOM anchors yet, so that is a real change rather
than a one-liner; and move detection, which is the bigger prize and feeds both channels
(verbatim moves dim, moved-and-edited highlight, with a nested diff of the move).

Stacked on #21.

🤖 Generated with Claude Code

https://claude.ai/code/session_018PkYQzbsnMihHesafWvXKs

[fiddur]

Pushed a fix and the test harness that should have caught it.

attentionClass and attentionTitle were declared on the props interfaces and used in the JSX but never destructured — a scripted edit looked for a two-line destructure where both files have a one-line one, and silently matched nothing. Every diff threw ReferenceError: attentionClass is not defined at render. An earlier variant of the same mistake put hunkAttention inside a useEffect closure.

Neither vite build (no typecheck) nor tsc stopped it, and there was nothing between a render-time ReferenceError and the browser. The UI package now runs component tests under jsdom — the React Router plugin is dropped for test runs, since its runtime preamble has no framework to come from, and JSX goes through esbuild instead.

Both new test files were written before the fix and reproduced the exact error. They assert the behaviour rather than that it renders: an imports-only hunk is dimmed and says why, real work is untouched, whitespace in a file where indentation is syntax is not dimmed, and a hunk the walkthrough points at is highlighted in preference to being dimmed.

164 UI tests pass.