Themes: Fix washed-out diff accents and invisible word highlighting

[masonmcelvain]

• close Theme derivation guards wash out diff colors and cause invisible highlighting #824

The two contrast guards in src/ui/themes.ts protected readability but destroyed theme identity in the process. This fixes both while keeping every existing contrast invariant.

Changes

• readableDiffSign now rescues low-contrast accents with the smallest black/white blend that clears the 3:1 floor, stepping up from 2%, instead of a fixed 45% wash.
  • Saturation loss is proportional to how far out of range the accent actually is. catppuccin-latte's green (contrast 2.96, missing the floor by 0.04) now gets a ~2% nudge (#40a02b → #3f9d2a) instead of being crushed to #235818.
• Word-emphasis tints (addedContentBg/removedContentBg) now derive first at full readable strength, and the row tints step further down until the pair clears a 12 channel-distance separation floor (readableSeparatedRowBackground).
  • This fixes invisible word-level highlighting on everforest-light, one-dark-pro, and material-theme-palenight (row and content had collapsed to identical colors) and near-invisible highlighting on plastic (distance 6).
  • Pushing content higher would break the readability guarantee that caused the collapse. This is a deliberate deviation from the issue's suggestion.

Tests

• Table-driven regression tests across all 66 bundled themes: accents that already meet the floor pass through unrescued, rescued accents stay within 2° of the source hue and clear the floor, and row-vs-content separation stays ≥ 12.

QA

• Spot-check in a real terminal: --theme catppuccin-latte (sign colors), --theme one-dark-pro or --theme everforest-light (word-level highlighting on changed rows).

bun run src/main.tsx -- diff --theme catppuccin-latte

Theme	0.19.0	Feature
catpuccin-latte (light rescue)
gruvbox-dark-hard (dark rescue)
snazzy-light (light rescue)
one-dark-pro (tint)
poimandres (tint)
everforest-light (rescue and tint)

[vercel]

@masonmcelvain is attempting to deploy a commit to the Modem Team on Vercel.

A member of the Team first needs to authorize it.

[greptile-apps]

PR author is not in the allowed authors list.

[benvinegar]

Pulled this onto fresh main and dug through it. The incremental sign-color rescue looks right, but I don’t think the PR is ready yet (feel free to push back on any of this).

The word-emphasis part seems to solve the same problem twice. diffRows.ts already checks whether a word-level highlight is distinct enough from its row background and strengthens it when needed, using a distance threshold of 28. This PR adds an earlier check in themes.ts with a threshold of 12. A color can therefore pass the new theme test and still be changed again by the renderer—77 of 130 current theme pairs do exactly that. Could we keep this policy in one place and test the final rendered highlight instead?

The new sign-color tests also don’t catch the old washout behavior. I replayed the old 45% blend against all 36 rescued catalog accents, and every one still passes the new contrast and hue assertions—blending toward black or white mostly preserves hue. A test for the first passing step, or at least a bound on how far the result can move from the source, would cover the regression better.

One smaller edge case: the rescue chooses white whenever background luminance is at most 0.45, but white can’t reach 3:1 once the background is above 0.30. For example, #aaaaaa ends up at about 2.32:1 against white, while black is over 9:1. None of the current bundled themes hit that range, but the helper should still uphold the contrast floor it promises.

I’d keep the incremental sign rescue, tighten its tests, and fold the emphasis fix into the existing renderer guard.

This comment was generated by Pi using GPT-5.6 Sol

[masonmcelvain]

Thanks for the review! I agree with all three points. I updated the PR comment "Feature" screenshots. The following response was written by Claude Fable 5:

One place I'd push back slightly: folding the emphasis fix into the renderer guard goes in the wrong direction. diffRows.ts can only strengthen emphasis color and it doesn't check text contrast — so when it rewrites a pair, it can quietly break the theme's 4.5:1 text guarantee. Only the theme layer can create separation by stepping the row tint down while keeping both surfaces readable. So I unified in the mirror direction: bundled themes now derive to the renderer's own floor through a shared constant, which makes the renderer guard a no-op for every bundled theme; the guard stays as a backstop for custom-theme overrides and transparent surfaces.

The test now asserts the final rendered highlight — resolveWordDiffHighlightBg must return the theme's pair untouched, with ≥28 separation and ≥4.5:1 text contrast — so the policy is one place and the tested color is the painted color. All 130 pairs still derive readable, visibly tinted rows at 28; the only value that moves is github-dark-default's added row, which steps down one 2% notch.

For the sign-color tests, I added the minimality assertion you suggested: each rescued accent must equal the first 2% blend step (from either anchor) that clears the floor. Replaying the old 45% wash against that fails for all 36 rescued accents.

And readableDiffSign now falls back to the opposite anchor whenever the luminance-chosen one can't itself reach 3:1, which closes the (0.30, 0.45] gap; since the fallback only fires when the preferred anchor can't clear the floor, every current bundled theme's output is byte-identical. Added a unit test with your #aaaaaa case.