INL1a 2/5 (B1): the inline input model (InlineBox, LineBreak, strut, verticalAlign) and its additive vector migration

[thejackshelton]

INL1a part B1: the R5 inline input model and its migration. Part 1 was #66.

Stacked, lands as train 3. This is the first of four stacked PRs (B1 → B2 → C1 → C2). They land as train 3, after train 2 (REPL-a). B1's head predates master's latest commits (it was built on master 30f03f4), so the train's catch-up merges master into it.

What changed

• Input model: InlineBox (kind inline), LineBreak (kind br), a required LayoutBox.strut (CSS2 §10.8.1: the container's own font and line-height when its children are inline-level, else null) and a required LayoutStyle.verticalAlign (keyword, px, percent or calc; the compiler writes baseline).
• Validator: validateLayoutInput checks the new shapes:
  • the strut is present exactly when the children are inline-level;
  • a text leaf's font and line-height must equal its parent's, the strut's or its inline box's (new code leaf-font);
  • no box or replaced leaf inside an inline box (block-in-inline);
  • display: inline only on inline boxes, so a replaced leaf stays block-level;
  • white-space collapsing is checked across inline boxes and <br>s.
• Engine: it still lays out text leaves only. box.ts inlineTextLeaves refuses an inline box or <br> with the typed code inline-box, naming the node. inline.ts is unchanged; B2 replaces the refusal with the inline core.
• Environment pass: zooms and resolves the strut, inline-box and <br> fonts, and vertical-align px and calc. boxDependencies covers them.
• Compiler and native emit: ios-layout.ts writes verticalAlign: baseline and the container strut (lowerFont, which refuses a non-Ahem family). Native emit writes LineStrut, InlineBox, LineBreak and VerticalAlignKeywordValue with the translated union names. native-program throws on an inline box or <br> rather than dropping it.
• Harness and corpus: the translate harness decodes the new kinds and fields, and the random corpus writes the strut.
• Migration: scripts/migrate-inline-input.ts migrates the hand-written inputs (17 calc vectors and the vectors README example), which no generator rewrites. It refuses to write anything unless the result differs only by strut and verticalAlign, and its --check is clean.
• Migration proof: docs/decisions.md requires a committed check that a vector migration is additive. node scripts/check-inline-input-migration.ts origin/master finds 4777 files at the base and 1921 changed, with 68602 neutral additions. Every other byte is identical, and no files are added. Its planted faults (vector-output, vertical-align, strut-font, strut-on-block, missing-strut, stray-file) each fail with their own message.
• Regenerated outputs (own commit, pnpm regen, fixed point after 3 passes): 476 vectors and 1428 DPR vectors gain strut and verticalAlign only. The other outputs are the 44 translated engine files, the corpus digests, the lanes.json host lanes and native-lanes.ts. No capture, emitted CSS, break, pixel or WPT output changed.

Tests and pins changed, with reasons

• parity.test dropInheritedText plant: it is now caught one step earlier. The reverted leaf font no longer equals its container's strut, so the validator rejects the input (leaf-font) before the layout lane runs. The case still fails, which is what the plant checks.
• native-backends.test: the expected constructor text gains verticalAlign: baseline as the last LayoutStyle argument. The calc line-height count goes from 2 to 3 because the strut adds one.
• README key-count test: it now compares against styleSchema's keys, so the new verticalAlign key is counted.
• New: packages/layout/test/inline-input.test.ts, with 10 tests covering the validator, the inline-box refusal and the environment zoom.

What passed (host-side)

• tsc -b passes.
• layout:subset reports 0 violations.
• parity.test passes 687/687.
• check-inline-input-migration.ts is clean, and all 6 planted faults are caught.
• pnpm test: 15 tests in 7 files fail, and every one of them needs device evidence (lanes, lanes-records, lanes-concurrent, device-failures, p6a-promotion and related). They pass or fail only after the train's device run and second pnpm regen, which rebuilds native-lanes.ts from the device-run lanes.json. No device run has been done for this head.

Size

About 98 KB of reviewed diff after .macroscope/ignore.md (25 files). Counting the hand-migrated calc vectors, which the ignore list skips, it is about 136 KB.

🤖 Generated with Claude Code