OVFL A 1/2: scroll containers in the engine (overflow auto, scroll, clip; Blink scrollable overflow port)

[thejackshelton]

OVFL Phase A, part 1 of 2 (T078, binding spec notes/T078-ovfl-spec.md): scroll containers in the layout engine and the translator. Stacked on INL1a (#88–#91); the review base is review/inl1a-lowering (afec4d6, the INL1a top). It lands as the OVFL train after the PNT1 train, together with A2 and Phase B (R14).

What changed

• Engine values. Overflow gains clip, auto and scroll, and the validator accepts css-overflow-3 §3.1 computed pairs only. hidden, auto and scroll make a scroll container. clip does not: margins still collapse through it, and a flex item keeps its automatic minimum size.
• overflow.ts. A port of Blink 145.0.7632.6 ScrollableOverflowCalculator (scrollable_overflow_calculator.cc/.h, plus the in-flow bounds in box_fragment_builder.cc 241-366). It computes the scrollable overflow and client size of every scroll container and of the viewport, with no scrollbar gutter.
  • It reads the layout's boxes, so layout() output is byte-identical.
  • Three shapes are refused rather than guessed (OVFL-p): a percentage relative offset inside a scroll container, the same on the root, and a percentage height with collapsing end margins.
  • Planted engine faults gutterReserved and overflowIgnoresPadding (appended).
• R16. PLACED_LINE_FIELDS names every PlacedLine field the line pass accounts for. A test fails typecheck when the line type gains a field, and fails on a stub extra item kind, so a new line item kind (INL2's atomic inlines) can't be skipped silently. The translated engine can't inspect fields at runtime, so this guard replaces a runtime refusal (PM ruling).
• Translator. Overflow literals, the FAULT_KEYS append, scrollMetrics and scrollMetricsWithFaults roots, and the engine-overflow suite (3000 generated cases) appended after engine-inline. targets.ts declares the suite.
• Port registry (R12). docs/ports.json registers scrollable_overflow_calculator.cc, its .h and box_fragment_builder.cc as BSD ports. paint_layer_scrollable_area.cc (LGPL) is a class A reference: the basis is css-overflow-3 §2.2 plus Chrome observation. overflow.ts now cites it as "matches Chrome", not as a port. THIRD_PARTY_NOTICES.md is regenerated.
• Regenerated outputs are in their own commit (/tmp/heavy-lease.sh pnpm regen, fixed point after 3 passes):
  • the translated Swift and Kotlin engines;
  • the corpus digests, which move because the faults objects gain the two keys (results unchanged);
  • the host lanes.json.
  • Every capture, vector, profile row and pixel output is byte-identical.

What passed

• pnpm typecheck: pass.
• pnpm test (heavy lease): 3541 passed, 12 failed. All 12 are the device-record tests (lanes, lanes-records, device-failures, p6a-promotion, lanes-concurrent). They fail only because lanes.json has no fresh device run: device step pending.
  • A first run under machine load (load average 49) also hit pr-review and merge-train failures; both passed on the rerun.
• engine-overflow is 3000/3000 on Swift and Kotlin (host lanes.json).
• chrome-ports.test.ts passes, and so does the new overflow.test.ts. The R16 test fails when the guard is removed.

Changed tests and pins, with reasons

• native-dpr-swift and native-dpr-kotlin suite lists gain engine-overflow (the new suite).
• corpus-dpr.test.ts gains the suite count.
• chrome-ports.test.ts: the LGPL entry count goes from 20 to 21 for the new class A entry. R12 requires the entry. The file is outside the spec's allowed list, and the PM accepted it.

Pre-landing review fixes

• Replaced leaves count in the scrollable overflow. overflow.ts now indexes REPL-a's replaced leaves, and a replaced child of a scroll container counts the way Blink's calculator counts it:
  • it adds its border box (its content is never scrollable overflow);
  • it takes part in the in-flow bounds, with its margins and the end padding;
  • the margin read-back never collapses through it.
• Atomic inlines are refused by name. refuseLineLevelBoxes throws OverflowRefusal for a box or replaced child of an inline formatting context (INL2's atomic inlines), instead of the plain "no end margin strut" error (R16).
• Corpus. The engine-overflow suite now generates block-level replaced leaves (1004 of its 3000 cases), so the Swift and Kotlin translations run this code too.
• Tests. Three new overflow.test.ts tests fail without the fix: a 100x200 image in a 100x50 auto box scrolls 200 high; the in-flow bounds count the image's margins and the end padding; and the refusal.
• Base. REPL-a (repl-a 5a4c407) is merged forward through review/ovfl-a-base (86c15c9 = inl1a-lowering + repl-a), the new review base, so this PR still reviews OVFL only. That merge resolves four source conflicts by intent (computed-checks, elements, ios-layout, pixel-reference). It is not regenerated: it is a review base only, and this branch regenerates (fixed point after 3 passes).
• Chrome proof. The overflow-replaced fixture (in A2, OVFL A 2/2: scroll containers in the compiler, viewport propagation and Chrome scroll metrics #97) proves this against Chrome's scroll metrics.

Outside the spec

• packages/parity/test/p6a-dash-oracle.test.ts: the LayoutBox literal gains INL1a's strut field (merge of the REPL-a base).

🤖 Generated with Claude Code