Repository navigation
Check exact editor characters independently of IME decoration - #109
Conversation
…and modern engines
…add owned input contracts
…s and retained limits
|
Exact-head gate update: independently verified Android run 42026037454 completed successfully for |
|
Independent audit recorded for exact head |
moikdev
left a comment
There was a problem hiding this comment.
Both required independent full base-to-head audits are recorded for exact head 3f87bf7a4182b74ba16c52c113c67613e5f6bb41, base c4ca362e7f3efccc59eacc4fdcaa7184fe0507c1.
- Correctness/security and collector/test soundness: gpt-6.1-sol / high, independent, read-only, no delegation or companion consultation; no actionable findings within scope.
- UX/input/test coverage/evidence provenance: gpt-6.1-sol / high, independent and read-only; no actionable findings within scope, including all21 actual published PNGs.
Reviewers verified source/APK splits,296 public hashes,249 current source hashes,525 retained capture hashes and the bounded claims. They inspected retained tests/builds rather than executing them. Root separately reverified the unchanged clean checkout, current source/public Git bytes, successful exact-head local/android58146385516 and actual pull_request Android run42026037454.
Current scoped results remain23 positives/seven cohorts/108 restoration checks, two intended negatives/zero passes/29 checks, and13 final read-only preservation checks. System-IME failures, original pre-probe history uncertainty, zero-result Japanese limits, private-original/cross-boot and broader platform/parity exclusions remain explicit. The change corrects test contracts/fixtures; production code/assets are unchanged. Expected-head merge verification is the final step.
|
PR #109 merged as |
Native editor characters can remain exact while IME composition adds an underline to
EditableText. The old readiness helper compared complete annotation objects and could reject that legitimate state. This change compares every UTF-16 character, keeps composition diagnostics separate, and adds guarded native-input fixtures with matched old-assertion controls.Closes #108. Part of #69, which stays open. Production/build sources and all eleven packaged assets/notices are unchanged; every native run uses the frozen
fec0a4bfapplication. The original system-keyboard/old-image failures remain unresolved.Validation:
classes3.dex; positive-only runner receipts remain failed and calibration is separate.Controlled cases explicitly select a silent test-APK IME and drive a native InputConnection. Normal/200% observations cover API26 revision1/WebView58 at physical1080×1920/420dpi, and the dedicated API36 revision7/WebView133 tablet at physical1920×1200/240dpi with phone/landscape overrides. Enlarged cases are functional smoke. Japanese text retention uses a zero-result query and adds no nonempty-grid veto-sensitivity claim. Model/semantic setup, injected keys and direct dispatch carry no physical-typing credit.
The first probes omitted keyboard switch history; its original pre-probe identity is unverified. Final baselines follow the documented cleanup and verify four raw settings. One older modern control failed its cleanup deadline; its owner stayed held and a separate sixteen-check recovery restored the original state. Failed/insensitive cohorts keep their outcomes. See the report and checkpoint for binary/source splits, all84 attempts, 525 retained capture hashes and 21 selected raw frame sets.
Real app smoke evidence, same production binary. The first pair contrasts test rejection/acceptance; the ordinary app screen has no error banner and no UI repair is claimed.
This provides no full Android8 keyboard, OEM/physical, accessibility, whole-Store/Raycast, #89 or production-signing clearance.