[tmashara]

What / why

--computed only read the resting state, so a dead-CSS removal that dropped a hover colour or a focus ring passed silently (#51). Now, after the resting walk, links and controls (a[href], button, input, select, textarea, summary, [tabindex]) are read again in :hover and :focus-visible, subtrees included, and diffed like any other key:

Real output: the visual test's link (a:hover{color:red} a:focus-visible{outline:2px solid blue} on the left only) at /, and 812 links at /docs, --width 1280 --computed:

computed: 4 findings; dom identical on 2 of 2 route(s)
  body > nav > a[href="#x"]:focus-visible outline rgb(0, 0, 255) solid 2px→rgb(0, 95, 204) auto 1px  /@1280
  body > nav > a[href="#x"]:hover color rgb(255, 0, 0)→rgb(0, 0, 238)  /@1280
  body > nav > a[href="#x"]:hover border-top-color rgb(255, 0, 0)→rgb(0, 0, 238)  /@1280
  body > nav > a[href="#x"]:hover > span.icon border-top-color rgb(255, 0, 0)→rgb(0, 0, 238)  /@1280
  states: only the first 500 links and controls forced on /docs@1280 (+312)

(border-top-color follows currentColor, as it does for resting colour changes.)

The snapshot also reads outline and text-decoration-line (the usual focus ring and hover underline). These live in a new SNAP_PROPS; pixel mode's PROPS is untouched, so ElementDiff.changed and accepted fingerprints don't move.

How states are forced

• One CDP session per side (snapshotStates in locate.ts). The resting walk keeps the elements on a page global and returns their keys in order; Runtime.getProperties + DOM.requestNode turn them into nodeIds with no DOM mutation (works for --selector :scope roots too).
• Per state, CSS.forcePseudoState on all of them at once (["hover"], then ["focus","focus-visible"]; a new force replaces the last), then the in-page walk visits only those elements under key:hover / key:focus-visible, reusing visit (children …:hover > span, …:hover::before).
• The CDP session opens before the walk keeps the global, and finally deletes the global and detaches, which clears every forced state, before it resolves; the page is resting for whatever runs next. Runs after the screenshots and locate, so pixels never see a forced state.
• A state change matching the same element's resting change is dropped (collapseInherited), so a resting change isn't repeated per state. State keys don't count towards dom or unmatched.

What forcing all at once misses: rules relating two forced elements (ul:has(a:hover) a:not(:hover) dims siblings only when one is hovered), :hover on ancestors (forcing never sets it, so li:hover > ul menus aren't seen), :active, :focus-within, script hover handlers. Per-element forcing would catch the first but costs a style recalc each (~10 s at 2000 links). It never invents a finding.

Caps

• First 500 per page per side, document order; the rest are counted (computed.unchecked, max of the two sides) and named on a states: line.
• Elements with more than 50 descendants are skipped (a <main tabindex=-1> would re-walk the page per state).
• The states share the existing 15 s walk budget and 50k element limit; failure modes are unchanged.

Cost (M-series mac, one side; both sides run in parallel)

links	resting walk	with states
50	26 ms	55 ms
500	146 ms	406 ms
2000 (500 forced)	582 ms	828 ms

How tested

• tests/visual/computed.test.ts: hover colour one side drops → only a:hover color (+ currentColor border); a dropped focus ring → a:focus-visible outline; a:hover .icon → keyed a:hover > span.icon; identical states → nothing, and a resting text-decoration change isn't repeated per state; 501 links → unchecked 1, global gone and a resting re-read equals the one before.
• tests/unit/computed.test.ts: a one-sided state key isn't markup (dom, unmatched), the resting dedupe, the states: line.
• The new visual tests bite on behaviour, not just at import: with CSS.forcePseudoState stubbed out, or the state passes skipped, the hover, focus, child and cap tests fail; without the detach or the global's delete the cap test fails; without the resting dedupe the dedupe test fails. Full unit (354) and visual (385) tiers pass after rebasing on --aria (Add --aria: diff ARIA snapshots per route #93), including accepted.test.ts and the pixel-mode pins, unchanged. typecheck, lint, format:check clean.
• Docs: docs/output.md § Computed styles, README and --help flag rows, skill reference/SKILL wording on unshot states.

Fixes #51

🤖 Generated with Claude Code