Preserve host alignment and native state; add reproducible problem gallery

[CodeinScrubs]

Summary

This is a stacked follow-up to #162, based on fix/audit-regressions-20260926. It adds concrete source-preserving repairs and an evidence gallery, not a claim that every language/device/renderer is globally production-certified.

• Preserve host physical alignment in React/Vue and the optional DOM stylesheet.
• Analyze displayed light-DOM prose without deleting hidden metadata or splitting visibly continuous URLs; retain author nodes, source and restoration ownership.
• Replace two .NET greedy recognizers with bounded scanners and keep trailing combining marks in technical ranges.
• Refresh WPF inherited/style/resource baselines while retaining local bindings and restoration behavior.
• Defer UIKit editable mutations during marked-text composition, retaining managed state for retry.
• Separate deterministic unit work alarms from unchanged isolated performance budgets and retain failure reports in CI.
• Update four development-toolchain advisory overrides, without unrelated lockfile platform churn.
• Add nine actual Chromium before/after fixture pairs, source/HTML/evidence/hashes, reviewed original user examples, contribution guidance and an honest readiness/alternatives review.
• Protect complete ASCII boundary-query literals and additive LTR phrases in all five core implementations. Retain source offsets, per-paragraph policy, and LTR-only scoped no-op behavior.
• Use a controlled browser clock for two demo callback checks; exact source/direction/status checks and separate real-timer playground coverage remain. Clock API.

Reproducible visual evidence at 967201d

Nine-pair gallery · Engineering packet · Readiness and remaining gaps

Same logical Persian/English source	Real Chromium fixture capture
Before: first-strong LTR base
After: RTL reading order, physical-left alignment retained

The three new 390 px mobile cases are curated logical reconstructions inspired by five October 2 captures, not recovered ChatGPT source, edited screenshots or a modified ChatGPT client. Originals remain private. The Retrieval: case explicitly declares Persian paragraph intent because its default character majority is LTR. English formulas retain independent LTR bases. Host physical-left alignment stays unchanged.

Fresh local verification of the mobile delta

• Minimum Node 22.12.0 full pipeline: 27 files, 779 passing tests, two Windows-specific skips; 95.36% line coverage. The new 19 core/Markdown cases were rerun after adding Unicode-neighbor negatives.
• 90/90 browser checks passed: Chromium, Firefox and WebKit, one worker; includes nine new mobile checks.
• 40 Android core tests passed, zero failures/errors; rerun after adding Unicode-neighbor negatives.
• Packed Markdown-It 13.0.2 / 14.3.1 / 15.0.1 consumers: strict TypeScript and peer-resolution checks, 941 canonical cases and 11 host-structure cases per parser passed.
• 941 corpus cases and 94 native security fixtures; generated data, types, lint, package depth, 91 Markdown documents/241 local links, nine gallery pairs, all builds and Action probes passed.
• Unchanged isolated 2,000 ms alarms: dense inline forest 724.46 ms, dense streamed list 513.89 ms average on this Windows machine; not universal latency promises.
• Fresh full and production-only audits returned zero known advisories at the low threshold. GitHub still reports the unmerged default-branch fast-uri alert chore(deps): bump gradle/actions/wrapper-validation from 0b6dd653ba04f4f93bf581ec31e66cbd7dcb644d to ed408507eac070d1f99cc633dbcf757c94c7933a #18; this branch uses patched 3.1.8. The alert was not dismissed.
• Local Swift, .NET/WPF and Rust execution was unavailable for this newer delta; their actual current-head hosted platform results are reported below. Earlier independent review applies to earlier changes, not an independent review of these new mobile repairs.

Follow-up boundary hardening at c82907d

Added direct first/last-line geometry assertions for the terminal period/Persian question mark, without relaxing assertions or timeouts. Chromium and WebKit passed; one Firefox context-teardown run reported a compositor failure and is retained as a failure. The unchanged three-test Firefox rerun passed.

The new supplementary-plane negatives reproduced attached-query false positives in .NET regex and JVM lookbehind. Both now check the full neighboring scalar/code point for only the two new ASCII query forms. Prefix/suffix cases include 𝒜 and 𐒠 in all five core suites. The Kotlin regression failed before the fix and all 40 Android core tests passed after it. The 19 JavaScript/Markdown tests passed again. Current-head native/compiler/security gate results are reported below.

Current native integration boundary: a whole-control Android TextView or UIKit direction is not an independent-base Markdown renderer. Keep differing prose/formula/code blocks separate, and preserve authored physical alignment and original source.

The final source delta passed 20/20 CI jobs in run 36986883745, including the supplementary-plane negatives on all five core implementations and the punctuation geometry checks on all three browser engines. CodeQL run 36986880207 also completed successfully for all five languages: C#, JavaScript/TypeScript, Java/Kotlin, Swift and Rust. Both are on exact head c82907d725cf9c86cb28b8914c08649e00744d15. Use those conclusions for the current head; old green results are not substituted for new ones. Current browser tests and readiness snapshot include the added boundary assertions and remaining limits.

Hosted gates for 967201d

CI run 36982395852 passed 20/20 jobs and CodeQL run 36982391670 passed all five languages on 967201d. A documentation-only follow-up at bff2a52 also passed all 20 CI jobs in run 36983360624. These results do not validate the newer c82907d delta. CodeQL is explicitly dispatched because its PR trigger is restricted to a main base while this draft is stacked.

Earlier verification of b999435

CI run 36936425858 passed 20/20 jobs, including Android emulator, Apple package/iOS simulator, .NET8/WPF/sample, Rust on three OSes, browser/packed-consumer/quality, SBOM and performance jobs. CodeQL run 36936426301 passed five languages. These results do not validate the newer mobile changes.

Earlier local .NET 10 source-linked core/WPF validation passed 5,853 assertions and 941 corpus cases, but is not a substitute for the fresh .NET 8 job. Earlier Firefox repeat-run compositor failures were retained, not counted as passes; the new complete one-worker browser run above passed.

Release/rollout boundaries

This is unreleased 0.4.0 source, not a new registry artifact or a ChatGPT/Codex integration. No merge or registry publication was performed. Installing published npm 0.4.0 or Maven 0.1.2 does not install these source-branch repairs.

Apple independent paragraph restoration, native extended-grapheme parity, hostile tiny-chunk streams, native-speaker review, physical device/IME/accessibility tests and downstream pilots remain acceptance work. Native HTML/ICU/BidiFormatter and other tools can be preferable at their respective layers. Neither universal superiority nor zero defects is claimed.