SELD-R2 PR 1: :hover, :active, :focus and :focus-visible compiled as per-assignment interaction states (R3, R4, R5, R7, R13)

[thejackshelton]

SELD-R2 PR 1, the compiler part (T064, binding spec notes/T064-seld-r2-spec.md). Built on origin/master, independent of FORM-a; no stack.

What changed

• Selectors (css/selectors.ts, analysis/match.ts). :hover, :active, :focus and :focus-visible parse in any compound and position, including inside :is(), :where(), :not(), :has() and :nth-*(… of S). They match against an InteractionState (none by default, so every existing match is unchanged). :focus-within, :visited, :checked and the rest stay refused.
• R4, R5 and R7 (analysis/interaction.ts). For each case, the candidates are found to a fixed point. From them come four dimensions: hover chains, active chains, and focus values (pointer focus without focus-visible, keyboard focus with it). Every combination is resolved in full, and each distinct resolution is stored once: a :focus that changes nothing is the none state (R5). The states are a second level under the case's app assignment and never count against the app state table. More than 256 distinct reachable states in one assignment, or more than 4096 combinations, are refused with DRAGON_UNSUPPORTED_SELECTOR naming SELD-R2s. seld-r2a's combinedStateRefusal and its MAX_STATE_ASSIGNMENTS count are deleted, because combined states now compile. Each state records what Chrome's CSS.forcePseudoState forces to reach it (a real pointer's whole chains, plus focus), and the parity forced cases use that.
• R3, the web gate (emit/web-css.ts). One condition is emitted per combination a state stands for. Every condition that needs a :hover candidate goes inside @media (hover: hover), Tailwind's shape. A condition whose :hover candidates are all off is also emitted without them under not all and (hover: hover). So a tap on a phone never leaves a hover style, and desktop is unchanged.
• R13, the hit model (project.ts, HIT_MODELLED). This is fail-closed. In the parity lanes' native compile, every interaction rule is refused in a case that has a pointer-reachable state and a compiled paint fact the hit test does not model. The refusal is [ios]/[android] and names SELD-R2b. It runs after the value checks, so a value the profile already refuses does not count, and an overflow value computed from a refused partner (overflow-y: auto beside a refused overflow-x: hidden, computeOverflowPair) does not count either. Outside the lanes it does not run, because every interaction rule is already refused on native (below). (An earlier version of this body said R13 was dormant in enforce mode. That was wrong: it added 10 music-player errors through the computed overflow-y on html, naming SELD-R2b, which would not have cleared them. Both causes are fixed, see the review round below.)
• Native refusal until the runtime (nativeInteractionRefusals). iOS and Android have no interaction runtime until PR 3 and PR 4, and no profile row. So every rule that tests :hover, :active, :focus or :focus-visible is refused on each configured native target with DRAGON_UNSUPPORTED_SELECTOR [ios|android] :<pseudo> is not supported on <target> yet: the native interaction runtime arrives with SELD-R2 PR 3 and PR 4. Web compiles them. Only the parity lanes' compiles (compileFixture, nativeCompile) set the internal option interactionLanes, which keeps the states on native so the forced cases and hit vectors prove each state's resolution ahead of the runtime.
• No native profile row from lane-only cases. A case the lanes compile on native only through interactionLanes (a user's compile refuses its interaction rules there) proves no iOS or Android row: the internal record lists those targets (laneOnlyNative), and the parity pipeline gives such a case no iOS keys, as it does for the web-only fonts cases. So the interaction fixtures prove web rows only, and the native profiles, support counts and north star claim nothing for them until PR 3 and PR 4 lift the refusal. The lanes still compile and compare the forced cases and hit vectors on native. Caught by android-profile.test (every case proving a promoted android row compiles unblocked for a user), which failed in the previous job run before this fix.
• Transitions and animations in interaction rules stay refused. The transition and animation lists cascade in the none state only, and the web output writes them on the base classes. So a transition or animation declaration (longhand or shorthand) in a rule that tests an interaction pseudo-class is refused on every target, as it was before this PR, rather than silently dropped. A transition set outside the interaction rule compiles, and the browser runs it when a state rule changes the property.
• Plants: comboStateDropped, webHoverUngated and hitUnmodelledNotRefused are new. With interactionRuleDropped, all four are caught in packages/dragon/test/interaction.test.ts.
• Parity. seld-r2a's forced cases (<case>~ix<k>, run beside the counted cases through forced-pseudo.ts) are kept. New fixtures: interaction-active and interaction-combo (forced cases, ltr and rtl), reject-interaction-focus-within and reject-interaction-cap. reject-interaction-active is removed because :active now compiles.
• docs/decisions.md records R1 to R16 under "Runtime styles and animation" and amends "Hover and focus" to "a tap never hovers (principle 3); focus and activation follow Chrome traces".
• Reused branches. seld-r2a's part A (27ca07d41d) was cherry-picked by intent onto master, its conflicts resolved beside ANIM-b1 and INL, and then reworked per R4 to R9. seld-r2b (runtime and traces) and seld-r2c (cursor) belong to PRs 2 and 3.

Left for later PRs (not in this one)

• No device runtime yet. Interaction rules are refused on iOS and Android (above) until the runtime PR (rt-interaction.ts, the R7 table in state.ts, R16) and the native glue PR lift the refusal. The lanes compile the states on native through interactionLanes.
• FORM-a dependent pieces. Real focusables (button, range, a[href]), tabindex and the R9 display/touchConsumesTap clauses wait for FORM-a, so pointer- and keyboard-reachable focus is empty today and focus is reached only through forced cases. The interaction-focus fixture keeps seld-r2a's forced-only shape.
• PR 2 and later: cursor (seld-r2c), the @media (hover: hover) fold (R3's authored shape), the hover-only-reveal warning, the UA focus-visible ring refusal, and the interaction-relayout fixture. The relayout fixture needs app state and the runtime.

Evidence

• Chrome check before the regen (interaction fixtures only, live Chrome 145, runFixture). All four layout fixtures pass in both directions, forced cases included: hover 12 cases, focus 8, active 16, combo 10. All three rejects pass.
• R3 in Chrome. Dragon's web output for interaction-hover was loaded in a touch context (hasTouch, isMobile: (hover: hover) false). Forced :hover on a's chain and a real tap on tall leave both at their none-state styles. In the mouse context the hover style applies.

Review round (precomputed review of ee1ae75)

Each fix has a test in packages/dragon/test/interaction.test.ts that fails without it (checked by running the new tests with the source fixes reverted: all five failed).

• High, web-css.ts:211 / animations.ts:85: transition and animation declarations in interaction rules were silently dropped on web. Fixed by refusing them (interactionRefusals), as before this PR. Test: refuses transition and animation declarations in an interaction rule on every target pins .a:hover { width: 60px; transition: width 1s linear }, .a:hover { animation-play-state: paused }, :is(.a:focus) { animation: k 1s } and .a:active { transition-duration: 1s } as refused, and .a { transition: width 1s } .a:hover { width: 60px } as clean.
• Medium, web-css.ts:211: state rules replaced the color: inherit a child under an animated colour gets (T065 R9). Fixed: the state-rule diff passes the same under flag as the base rule, so an inherited colour stays inherit in every state, and a colour the child declares in a state is still emitted. Test: keep an inherited colour inherit under an animated colour in every state pins the reviewer's document.
• Medium, selectors.ts:159: native targets accepted the four pseudo-classes with no profile row and no runtime. Fixed by the native refusal above. Test: refuses every interaction rule on native outside the lanes pins the four [ios]/[android] errors for .a:active and .card:hover .title, blocked native targets, a ready web output, and no errors for web alone, for the lanes and for a document without interaction rules.
• Follow-on (PM, from the job run): native rows promoted from lane-only cases. Fixed as above (laneOnlyNative). Tests: android-profile.test after the regen, and the laneOnlyNative pins in refuses every interaction rule on native outside the lanes.
• Low, project.ts:687: R13 counted an overflow value computed from a refused partner, and the body said R13 was dormant. Fixed (above) and the body corrected. Test: counts an overflow value computed from a refused partner as refused, and leaves R13 to the lanes.
• Low, interaction.ts:84: isFocusable differed from HTML. An invalid tabindex (HTML's rules for parsing integers) no longer makes an element focusable, and input type=hidden never is. Test: focusability (R9). R9's general display: none clause still waits for FORM-a (PR 2): known follow-up.
• Low, cases.ts:36: Chrome never renders the conditions of collapsed combinations past the first member, nor the not all and (hover: hover) branch. Known follow-up, not fixed here: PR 2 brings seld-r2c's exclusivity and identity checks over the emitted CSS, and the touch-context Chrome run belongs with the R2 trace lane (PR 3). Today these are covered by the structural tests in interaction.test.ts and the manual touch-context check under Evidence.

Outside the spec

• packages/parity/test/glyph-clearance-pins.json: new case lines only (DRAGON_PIN_WRITE=1), additive.
• packages/dragon/test/compile.test.ts, selectors.test.ts, packages/parity/test/css-escapes.test.ts, device-vectors.test.ts and packages/translate/test/translate.test.ts: retargets listed below.
• packages/parity/src/pipeline.ts (compileFixture) and packages/parity/src/native-host.ts (nativeCompile): set interactionLanes, so the lanes keep the interaction states on native.

Changed tests, each with its reason

• interaction.test.ts: its compile helper compiles as the lanes do (interactionLanes) by default, because those tests prove the states on native as the lanes use them; the new tests pass lanes = false to pin the user-facing native refusal. No existing expectation changed.
• selectors.test, compile.test: the refusal pins move from :active, :hover and :focus to :checked, :visited and :focus-within, which stay refused. Intent kept: interaction and state pseudo-classes outside SELD-R2 are refused.
• css-escapes.test: div:\68over is now judged against Chrome (it compiles), and the escaped :focus-within refusal is pinned instead (selectors 59 to 60).
• interaction-forced.test: the registry lists the four forced fixtures, and the state keys are the new match-set keys. It adds a pin that a hover-and-press combination forces both chains.
• device-vectors.test, translate.test: p1/hit goes from 2028 to 2060, one hit vector per layout case per DPR (4 fixtures × 2 directions × 4 DPRs). The floor in p1-floor.json is unchanged.
• parity.test, platform.test, values.test, fixture-reader.test, android-profile.test, pixel-reference.test: seld-r2a's retargets. Forced cases are counted beside the reachable cases and pass the same prepare hook to both renderings; the values group order ends with interaction; reject-selector-hover refuses :focus-within.
• reject-interaction-active fixture: removed, because :active compiles (R4). reject-interaction-focus-within takes its place as the interaction-pseudo refusal.

What passed (job seld-r2-c2, at 3991a30)

• pnpm regen reached its fixed point; its outputs are in their own commit.
• pnpm typecheck: pass.
• pnpm test: Tests 15 failed | 4812 passed | 2 skipped (4829). Failing files: interaction.test.ts device-failures.test.ts land.test.ts lanes-concurrent.test.ts lanes-records.test.ts lanes.test.ts p6a-promotion.test.ts . The device-record files are device step pending (land.test only its evidence:stamp case); rerun alone and passed: interaction.test.ts.
• vitest run packages/parity/test/android-profile.test.ts alone after the regen: pass.
• Targeted before the job: vitest run packages/dragon/test/interaction.test.ts compile.test.ts selectors.test.ts packages/parity/test/interaction-forced.test.ts: 154 passed. The five new review tests fail with the source fixes reverted.

Music player north star (examples/music-player/dragon/north-star-check.json, against master)

• errors 356 -> 362
• web DRAGON_UNSUPPORTED_SELECTOR: 9 -> 4
• web DRAGON_UNSUPPORTED_VALUE: 43 -> 44
• ios DRAGON_UNSUPPORTED_SELECTOR: 9 -> 8
• ios DRAGON_UNSUPPORTED_VALUE: 52 -> 53
• android DRAGON_UNSUPPORTED_SELECTOR: 9 -> 8
• android DRAGON_UNSUPPORTED_VALUE: 52 -> 53

The native DRAGON_UNSUPPORTED_SELECTOR counts are the native refusal of the interaction rules until PR 3 and PR 4 (owned rises, cleared by those PRs); R13 adds none outside the lanes.

Reviewed size: 188438 bytes (Macroscope ignore patterns excluded), over the ~150 KB guide by the seld-r2a carry-over and the interaction tests; one theme, so it is not split.

🤖 Generated with Claude Code

[thejackshelton]

Landing stopped at step merge (pnpm land)

merge-train: merging seld-r2-c onto c78674a614f39a3ed5ee30aca2f68051b45a059d failed; conflicts in packages/dragon/src/faults.ts, packages/dragon/src/project.ts, packages/parity/src/fixtures.ts, packages/parity/test/device-vectors.test.ts, packages/parity/test/values.test.ts, packages/translate/test/translate.test.ts

Fix the cause, then hand #136 back to the landing queue with the new clean head.