INL2a 1/2: atomic inlines in the layout engine (inline-block, inline-flex)

[thejackshelton]

INL2a part 1 of 2 (T059, notes/T044-inl-spec.md R1-R8, notes/T059J-inl2.md): the layout engine for atomic inlines (inline-block and inline-flex). This is the INL2 chain restacked onto the v2 text stack: inl2a-engine → inl2 → txt2a-engine → txt2a → inl2b, as -v2 branches. The originals are untouched.

Base: review/tdec-a2-v2, a copy of tdec-a2-v2 7b99e3a, the top of the text-stack chain. This PR lands after the text-stack train.

What changed

• Engine (packages/layout), cherry-picked from inl2a-engine 62cad4640f:
  • An inline-block or inline-flex box is a LayoutBox among a block container's inline-level children.
  • inline-box.ts lays it out at its shrink-to-fit width (CSS2 §10.3.9), with its margin box as its advance.
  • Its baseline is the last line box's for inline-block (the bottom margin edge when there is none, or for a scroll container), and the first baseline for inline-flex.
  • inline.ts puts it in the break text as U+FFFC with a soft wrap opportunity on each side, places it on its line, and counts its min- and max-content contributions.
  • block.ts carries the last baseline through block children.
• Refused, with typed codes: an atomic inline beside text in a face other than Ahem, rtl without a letter on both sides, relative positioning, a percentage block size, and an atomic inline inside an inline box (validate.ts).
• Six engine plants. The engine-inline differential suite appends 1500 atomic contexts after INL1a's 3000. inline-atomic.test.ts compares the 16 family-5 cases with Chrome at DPR 1, 2, 3 and 2.625, in ltr and rtl.
• Merge with the new base (intent-based):
  • REPL-a's block-level replaced boxes are now in the base. A replaced child gives an inline-block no last baseline, because it has no line boxes. The block.ts and flex.ts replaced paths set lastBaseline: null.
  • inlineChildren still rejects a replaced box beside inline content.
  • validate.ts checks for an atomic inline first, then applies the base's isInlineLevel block-in-inline rule.
• docs/ports.json (the port registry from PORT-0: Chrome port registry (docs/ports.json), its test, and generated THIRD_PARTY_NOTICES.md #52 postdates the old base), with notices regenerated by pnpm notices:gen:
  • block_layout_algorithm.cc, flex_layout_algorithm.cc and line_breaker.cc gain their Dragon references;
  • physical_box_fragment.h, inline_items_builder.cc and logical_line_builder.cc are new BSD reference entries.
• Changed test: corpus-dpr.test's engine-inline count is derived from INLINE_SPEC, now 3000 plus the atomic contexts. The base's text-latin count is kept.

What passed

• pnpm regen: fixed point after 2 passes. The output commit covers the translated engine, corpus-dpr.json, host lanes and profiles.
• pnpm typecheck: clean.
• pnpm test (heavy lease, on a loaded machine): 178 files; 3723 passed, 39 failed and 1 skipped.
  • 7 in chrome-ports: the missing registry entries, fixed in 6518c6f. The file now passes 13/13.
  • Device step pending: lanes, lanes-records, lanes-concurrent, device-failures and p6a-promotion. lanes.json is host-only, and the device run happens in the train.
  • The rest were load timeouts. All pass on rerun; inline.test passes 12/12 when run alone.
• North star: unchanged by this PR, because the engine alone changes no compiler output. The INL2 delta is reported on the inl2 PR.
• Macroscope is at its spending limit, so this PR relies on CI.

🤖 Generated with Claude Code