Rebase and re-audit the accessibility core candidate on DSH 0.1.2-alpha.2

[mattheliu]

Outcome

Produce a reviewable accessibility-core candidate for the newly published DSH 0.1.2-alpha.2 line without treating the completed alpha.1 evidence as transferable conformance evidence.

Observed boundary

• Official npm alpha: @deepseek-ai/dsh@0.1.2-alpha.2.
• Upstream tag and current master: dsh-v0.1.2-alpha.2 / 0a53fb55bea101816fa226bb964ae2bed71c343b.
• The alpha.1-to-alpha.2 delta changes 1,604 paths.
• It overlaps 128 of the 302 paths changed by alpha.1 candidate d809463d916b05acc3d6fc5821e83832e1d5e0c8.
• The alpha.1 candidate and its attempt-1 19/19 CI run remain immutable evidence for alpha.1 only.

Work

• Derive a fresh portability base from the exact alpha.2 tag.
• Audit all 128 overlapping paths as upstream-equivalent, conflict, reapply, redesign, or drop.
• Re-audit new or reorganized core surfaces outside the overlap for semantic, keyboard, focus, announcement, reflow, contrast, motion, error-recovery, and safety barriers.
• Port only core-owned behavior that remains necessary.
• Reconcile browser-state and coverage-owner boundaries with alpha.2 instead of copying alpha.1 test scheduling mechanically.
• Preserve every failed checkpoint without rerunning it to hide failures.
• Update the bilingual compatibility ledger and public review thread.

Acceptance evidence

• Exact alpha.2 base and candidate commits are recorded.
• The 128-path overlap inventory and every disposition are public.
• Type, unit, build, package, assembled-browser, keyboard, Chromium/Firefox/WebKit accessibility, Linux coverage, Windows coverage, and native Windows gates pass at an exact checkpoint.
• Automated evidence is kept distinct from VoiceOver, NVDA, braille, target-system, and disabled-developer task evidence.
• No companion peer range, npm release, tag, or complete-accessibility claim is authorized until its separate gates pass.

[mattheliu]

Alpha.2 core migration checkpoint is now public for review: omdsh-dev/deepseek-harness#2

Version boundary:

• official baseline: dsh-v0.1.2-alpha.2 / 0a53fb55bea101816fa226bb964ae2bed71c343b
• candidate head: 7615cb8e2c25a6f413162d6df9a3142c0cb4fc05
• validation portability: c794ffcd1c
• modal and Workspace accessibility slice: 409564d45e
• fork standard-runner policy: 7615cb8e2c

Implemented in this slice: shared modal application inertness, initial/contained focus, nested topmost dismissal, descriptions, connected focus restoration, native Button refs, and the Workspace adoption-error focus/description path. Fork PR jobs now select standard ubuntu-24.04 and windows-2025 runners with bounded four-core workloads; live run 33324826327 confirms those labels, while upstream-only preview and Project mutation are neutral skips.

Local evidence: component 50/50; full GUI 3,859 pass / 4 skip; assembled Workspace 13/13; workflow contract 20/20; Issue policy 23/23; doc-sync 32/32; lint, typecheck, full build, pre-commit, and pre-push passed. The first full Web replay remains recorded as 89 files pass / 1 skip / 2 fail and 307 tests pass / 16 skip / 1 fail; after deterministic Node 24.3 HMR and remote-host fixes, the focused HMR, remote-welcome, and dev-web regressions passed. It is intentionally not restated as a green full replay.

Current CI is pending: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33324826327

This remains a partial alpha.2 candidate. No complete-accessibility claim, merge, tag, or npm publication is authorized by this checkpoint. Real VoiceOver/NVDA/other AT sessions, disabled-developer evidence, remaining interaction owners, and a dedicated cross-browser accessibility gate remain open.

[mattheliu]

Alpha.2 correction checkpoint (public evidence, no release claim)

• Exact head: 7546b3ba5541c89d59bb616e4b1eafe9e154c3bc
• Preserved failed CI: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33324826327
  • Linux coverage: the persistent PowerShell test observed the supported quiescent inferred_idle tier instead of prompt-observed stdin_read; the following send/state/secret assertions were not weakened.
  • Snapshot lane: the two PowerShell header-owner recordings were stale after alpha.2 runtime-context composition.
  • Windows coverage completed successfully; no failed job was rerun.
• Deterministic correction:
  • accept either supported ready tier for the first persistent-PowerShell send;
  • refresh the session, system-prompt, and tool-schema owners together using a SHA-256-verified PowerShell 7.6.5 portable runtime;
  • record the evidence rule in the bilingual alpha.2 Agent Note.
• Local evidence at the exact head:
  • real PowerShell file: 8 passed, 1 Linux-only skip;
  • focused PowerShell header owners: 2 passed, 77 out-of-scope skipped;
  • doc-sync: 32/32;
  • typecheck and staged/full lint: pass;
  • pre-push typecheck: pass.
• Preserved limitation: the complete macOS snapshot replay additionally exposed seven ACP startup-order mismatches where a duplicate config_option_update precedes session/new; the same ACP cases passed in the Linux CI attempt. This is not counted as green and is not folded into the PowerShell expectation.
• New, non-rerun CI: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33326266154

This remains an alpha.2 migration candidate. It is not a claim that DSH is fully accessible, and it is not approval to merge, tag, or publish. Real VoiceOver/NVDA/other assistive-technology task evidence and disabled-developer validation remain required.

[mattheliu]

Alpha.2 evidence update: second run preserved, deterministic corrections pushed

Second CI run (attempt 1, no reruns): https://github.com/omdsh-dev/deepseek-harness/actions/runs/33326266154 at 7546b3ba5541c89d59bb616e4b1eafe9e154c3bc. The failed run remains part of the evidence history.

Observed failures and corrections:

• Linux coverage exposed two noncanonical packed PowerShell session fixtures and a real PowerShell startup race. The fixtures were rewritten with the repository migration command. PowerShell startup now observes native startup output without input, submits the controlled-prompt bootstrap exactly once, requires both backend stdin_read and the owned OSC marker, and returns only the concise controlled prompt instead of echoed bootstrap source.
• Windows coverage exposed the same packed fixtures plus two projection-cache tests that assumed a fixed 40 ms sleep drained fire-and-forget filesystem writes. Those tests now wait, within a bound, for the exact stored checkpoint or trigger-specific warning; the injected failure and self-healing assertions remain unchanged.
• Windows native tests preserved an unexpected Vitest worker-fork exit after four peer files passed. The complete 54-test worker-thread file now runs as four named suites in four isolated Vitest processes, followed by the remaining native specs. No assertion was removed and no retry was added.

Local evidence on the corrected tree included:

• real PowerShell terminal integration: 8 passed, 1 host-inapplicable skip;
• both PowerShell snapshot replays through the shipped headless profile: 2 passed;
• terminal startup and packed-layout focused tests: 26 passed;
• projection-cache regression file: 17 passed;
• four isolated worker-thread suite processes: all 54 tests passed;
• CI workflow contract: 20 passed;
• documentation/type/export/i18n gates: 32 passed;
• staged pre-commit and pre-push/typecheck hooks: passed.

Correction commit: 3cc585a2d4f9cab37bc7660798d05aac3d7da2b0
Third CI run (attempt 1): https://github.com/omdsh-dev/deepseek-harness/actions/runs/33327867283

This is migration evidence, not a release gate completion. No merge, tag, npm publication, or whole-product accessibility claim is being made. Real VoiceOver/NVDA and disabled-developer task evidence remain required.

[mattheliu]

Alpha.2 keyboard and browser contract checkpoint

Candidate head: 2359d51d29351a777a4af013218b7f30f7063ddd

• PR: feat(a11y): add core keyboard and browser contracts deepseek-harness#2
• Preserved failed run (not rerun): https://github.com/omdsh-dev/deepseek-harness/actions/runs/33327867283
• New attempt-1 CI: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33330458425

Implemented in this checkpoint:

• APG-style roving Workspace tree behavior across grouped, flat and search views, including arrows, Home/End, expand/collapse, parent/child movement, activation, typeahead and focusable row actions.
• Shared Settings dialog ownership for application inertness, contained focus, topmost Escape dismissal and connected focus restoration.
• Persistent PowerShell readiness now requires an observed return to stdin wait; same-shell echo plus silence cannot produce inferred-idle readiness.
• A required assembled-app Chromium, Firefox and WebKit accessibility job on standard GitHub-hosted fork runners.

Local evidence on the exact head:

• Workspace tree owner: 52/52 tests, 100% statements/branches/functions/lines.
• Terminal readiness: 59 passed, 3 local no-pwsh capability skips, 100% statements/branches/functions/lines. Those skips are not PowerShell evidence.
• Built assembled app: 2/2 paths passed independently in Chromium, Firefox and WebKit.
• Documentation quick gate 15/15 and full synchronization 32/32.
• Full production build, lint, typecheck, pre-commit and pre-push hooks passed.

Boundary: this remains an alpha.2 migration checkpoint, not a whole-product accessibility or release claim. Automated browser contracts do not replace task-level VoiceOver/NVDA/other assistive-technology, braille/input, or disabled-developer evidence. The new CI is still running; no merge, tag or npm publication is proposed.

[mattheliu]

Alpha.2 accessibility-core candidate advanced to dc9cc8d6145a7007c8f23a987234daa7f2442d09 in omdsh-dev/deepseek-harness#2.

Preserved run 33330458425 was not rerun. Its concrete failures were analyzed and corrected: exported keyboard-hook documentation, the pnpm-only accessibility-browser invocation, alpha.2 ARIA-level/empty-tree snapshot expectations, fast persistent-PowerShell stdin-wait evidence, and the Windows-instrumented SQLite busy-retry fixture budget.

Local corrective evidence: terminal session 54/54 with 100% per-file coverage; SQLite schema owner 86/86; recorded-session snapshots 114 passed with 2 capability skips; full Web replay 303 passed with 15 capability skips; CI static 43/43. The macOS host has no pwsh, so the real-PowerShell composition skip is explicitly not passing evidence.

New attempt-1 main CI: https://github.com/omdsh-dev/deepseek-harness/actions/runs/33332020842