TXT1a-1a (v2): Ahem through HarfBuzz, and the zoomInput platform rule, restacked onto INL1a C2

[thejackshelton]

Stacked, lands as the text-stack train after train 3. This branch is a restack onto INL1a (#88 B1, #89 B2, #90 C1, #91 C2), made by intent: C2 plus this branch's own change. Its base is review/inl1a-lowering, pushed at inl1a-lowering's head afec4d6, so the diff is this branch's alone. The text stack lands as one train after train 3; master is merged in during the train's catch-up. Host-only: no device run yet (pending-device).

What changed

• TXT1a-1 phase A (T083J), restacked from txt1a-1a (4834ada88): Ahem is shaped through HarfBuzz on the host (text-shaper-host.ts, shaping.ts), and a planted platform rule resolves the environment through zoomInput, which keeps zoomInput reached in the translated engine.
• Applied without conflicts.
• docs/ports.json: block.ts (the noReshapeAtBreak plant cites shaping_line_breaker.cc) and text-shaper-host.ts hanKerningOf (han_kerning.cc) are added as references. PORT-0 landed on master after the branch was written. THIRD_PARTY_NOTICES.md does not change.
• Regenerated outputs: only the translated engines, the corpus digests and the lanes.json host lanes change. Every capture, vector, break, pixel and profile is byte-identical.

Tests and pins changed, with reasons

• None beyond the branch's own new test (text-latin.test.ts phase A).

What passed earlier (host-side, superseded head c0d4cd2, before the cloud migration)

• /tmp/heavy-lease.sh pnpm regen reaches a fixed point; its outputs are in their own commit, whose message names the command.
• pnpm typecheck passes.
• pnpm test: everything passes except the 12 device-record tests that need a fresh device run (lanes, lanes-records, lanes-concurrent, device-failures, p6a-promotion): device step pending. The first full run failed chrome-ports.test.ts on the two citations above. They are fixed in this branch, and chrome-ports.test.ts passes (13/13).

Catch-up with master (2026-10-08, land-prep)

• git merge origin/master (157 commits, including INL1a 5/5 (C2): lowering inline boxes and <br>s, native runtime emit, 7 inline fixtures and lanes #91 INL1a C2, the cloud migration, OVFL A: scroll containers (overflow auto, scroll, clip), viewport propagation and Chrome scroll metrics, rebuilt on master (supersedes #96/#97) #194 and CASC 1: @supports decided at build time as Chrome 145 decides it; revert and revert-layer; CSS-wide keyword rows #200) as merge commit 6c5575f. Only one file conflicted as text:
  • packages/translate/harness/harness.ts FAULT_KEYS: this branch adds the eight text faults (advanceNot16_16 ... latinCheckSkipped) and master adds gutterReserved, overflowIgnoresPadding. Resolved as the union, every key from both sides kept, in the existing order (branch keys, then orderHalfEven, orderUnclamped, then master's two). decodeFaults auto-merged with all keys from both sides.
• Generated outputs (merge=dragon-generated) kept this side and were rebuilt by regen on CI. No floor or pin file conflicted.
• pnpm install --frozen-lockfile was clean.

Tests changed at the catch-up

• None. No test, check, tolerance, floor or fixture changed or was removed.

Review round 1 (independent review at 0c6b3ee, two Medium findings), commit 76150ba

1. latinCheckSkipped was inert, and the harness ran shaping plants inert.
   • R4 is now implemented on the engine's shaped measurer: shapedMeasurerFor wraps the measurer in latinScopedMeasurer (shaping.ts), which refuses text with any code point outside Latin, Common and Inherited (isLatinText, script-data.ts) with a new typed code, text-script (unsupported.ts, section notes/T056-txt1a-spec.md R4). latinCheckSkipped skips exactly that check. The shaping core (shapedText) stays script-agnostic, because the TXT1-S gate shapes every script. This is observable on Ahem: the bundled Ahem.ttf has glyphs for 45 non-Latin code points (for example U+03A9 and U+6C34), which shaped Ahem measured before this change.
   • So the engine can raise the typed code, MeasureResult failures now carry code: 'text-glyph' | 'text-script', and inline.ts raises each code as a literal (unsupported.test.ts requires every code to be raised literally).
   • The translate harness (runEngineCase) now reports a harness error for any shaping plant (advanceNot16_16 through latinCheckSkipped). It lays out with measurerFor's Ahem measurer, so before this change those plants did nothing there. No corpus line sets one (the corpus fault lists are fixed and exclude them), so the corpus outputs are unchanged.
2. Only metricRoundingSwapped was planted in text-latin.test.ts. The header now states exactly what is proven, and each shaping plant has evidence:
   • wholePixelPositions changes the committed text-fractional-font-size vector.
   • advanceNot16_16 and doubleAccumulation change no committed vector (measured: 0 of the committed vectors). They do change widths in T082's committed Chrome runs (7.77px, 93 or more glyphs), and they are asserted to miss Chrome there.
   • kerningDropped, noReshapeAtBreak and softHyphenWidthMissing cannot act on Ahem through the engine's measurer, and are pinned inert on every committed vector with the evidence asserted:
     • Ahem.ttf has no GSUB, GPOS, kern, morx or kerx table, so kerning and reshaping change nothing.
     • noReshapeAtBreak and softHyphenWidthMissing act only in ShapedText.line, which TextMeasurer does not expose.
     • No committed vector has a U+00AD.
     • shaping-gate.test.ts still catches all three on the real-font gate.

Tests changed, with reasons (review round 1)

• packages/parity/test/text-latin.test.ts: four tests added (wholePixelPositions; advanceNot16_16 and doubleAccumulation against Chrome; the three inert plants with evidence; R4 text-script and the latinCheckSkipped plant, which changes the outcome to the line breaker's later refusal and makes the measurer return ok) and one inert guard for latinCheckSkipped on the committed vectors. The header claim is narrowed to what is proven. The R4 test fails without the fix.
• packages/layout/test/inline.test.ts: added a test that a measurer's text-script refusal reaches the layout result with its own code and section. It also asserts the section of the existing text-glyph refusal.
• packages/layout/test/font-data.test.ts: the old-measurer stub's refusal carries code: 'text-glyph', because MeasureResult gained the field. The assertions are unchanged.
• packages/translate/test/translate.test.ts: added a test that the harness reports a harness error naming each of the 8 shaping plants, and lays out the same input without one.
• No tolerance, floor, pin or existing check was loosened or removed.

What passed on head 09fff4c

• pnpm ci:regen txt1a-1a-v2 --head 76150ba95a1d138afa520caee0cb899b4a5b2df4 --once: regen commit 09fff4c, https://github.com/compiled-run/dragoncss/actions/runs/37858585385. The earlier catch-up regen (6c5575f -> 0c6b3ee) was https://github.com/compiled-run/dragoncss/actions/runs/37849014715.
• pnpm typecheck (local): passes.
• Platform-free, local: packages/layout/test/{font-data,inline,vectors,dpr-vectors}.test.ts and packages/dragon/test/{registry-claims,profile-index}.test.ts: 6 files, 762 tests passed.
• pnpm ci:test-files 09fff4cafd... --once, https://github.com/compiled-run/dragoncss/actions/runs/37862735384 (tested 09fff4cafdaff3a6912c0d2318622d9e8e0ccfb5), all passed:
  • Chrome group: translate 16, android-profile 5, chrome-ports 13, dist 6, env-engine 8, fixture-reader 7, fonts-fixtures 26, parity inline 13, interaction-forced 3, media-runtime 8, media-sweep 19, p6a-dash-oracle 310, parity 518, phrasing-blockified 3, platform 12, parity registry-claims 8, replaced-identity 3, text-latin 10, tree-faults 3, values 72.
  • Native group: native-swift, native-kotlin, native-dpr-swift and native-dpr-kotlin.
• pnpm -s pr:review 101 --once: exit 0 (CI checks success; 0 unanswered findings; UNREVIEWED, because Macroscope is at its monthly spending limit).
• Device lanes: device step pending (the landing driver runs them).

🤖 Generated with Claude Code

---

Generated by Claude Code