test: add accessibility and Python 3.13 evidence

.github/workflows/ci.yml

@@ -38,6 +38,74 @@ jobs:
           (cd "$RUNNER_TEMP" && scopeproof-web --version)
           (cd "$RUNNER_TEMP" && python -m scopeproof_core.evals.r002_swebench --help)
 
+  compatibility-python-313:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
+        with:
+          python-version: "3.13"
+          cache: pip
+      - name: Upgrade packaging tools
+        run: python -m pip install --upgrade pip
+      - name: Install
+        run: python -m pip install -e '.[dev,research]'
+      - name: Test
+        run: python -m pytest -q
+      - name: Deterministic benchmarks
+        run: |
+          python -m scopeproof_core.evals.runner
+          scopeproof comparison-benchmark
+      - name: Installed wheel compatibility smoke
+        run: |
+          wheel_dir="$RUNNER_TEMP/scopeproof-wheel-python-313"
+          python -m pip wheel . --no-deps --wheel-dir "$wheel_dir"
+          wheel="$(find "$wheel_dir" -type f -name 'scopeproof-*.whl' -print -quit)"
+          python -m pip install --force-reinstall --no-deps "$wheel"
+          python -m pip check
+          (cd "$RUNNER_TEMP" && python -c 'from importlib.metadata import version; from scopeproof_core import __version__; assert version("scopeproof") == __version__')
+          (cd "$RUNNER_TEMP" && scopeproof --version)
+          (cd "$RUNNER_TEMP" && scopeproof-web --version)
+          (cd "$RUNNER_TEMP" && scopeproof benchmark)
+          (cd "$RUNNER_TEMP" && scopeproof comparison-benchmark)
+
+          web_log="$RUNNER_TEMP/scopeproof-web-python-313.log"
+          web_pid=""
+          cleanup() {
+            if [ -n "$web_pid" ]; then
+              kill -- -"$web_pid" 2>/dev/null || true
+              wait "$web_pid" 2>/dev/null || true
+            fi
+          }
+          trap cleanup EXIT
+
+          setsid scopeproof-web --host 127.0.0.1 --port 8513 > "$web_log" 2>&1 &
+          web_pid=$!
+
+          healthy=0
+          for attempt in $(seq 1 30); do
+            if ! kill -0 "$web_pid" 2>/dev/null; then
+              echo "Installed Python 3.13 workbench exited before becoming healthy."
+              cat "$web_log"
+              exit 1
+            fi
+            if response="$(curl --fail --silent http://127.0.0.1:8513/_stcore/health 2>/dev/null)"; then
+              if [ "$response" = "ok" ]; then
+                healthy=1
+                break
+              fi
+            fi
+            sleep 1
+          done
+
+          if [ "$healthy" -ne 1 ]; then
+            echo "Installed Python 3.13 workbench did not become healthy within 30 seconds."
+            cat "$web_log"
+            exit 1
+          fi
+
   locked-environment:
     runs-on: ubuntu-latest
     steps:
@@ -59,7 +127,7 @@ jobs:
         run: python -m uv run scopeproof comparison-benchmark
 
   verify:
-    needs: [compatibility-python-311, locked-environment]
+    needs: [compatibility-python-311, compatibility-python-313, locked-environment]
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0

docs/audits/accessibility-platform-evidence/verification.md

@@ -0,0 +1,134 @@
+# Accessibility and platform evidence verification
+
+## Evidence boundary
+
+- Date: 2026-08-09 (America/Toronto).
+- Branch: `codex/accessibility-platform-evidence`.
+- Final executable/test verification HEAD: `1a3defcfec5f9f2231a67e8f0a4450d43a63b9b4`
+  with tree `b01e358b5aefefd7eaf8abf28e0bf7502cf962af`.
+- Base: `origin/main` at `30177733ef312ced22e6a2e57e3df6fdb1e92507`.
+- Host: macOS 26.5.1 build 25F80 on Apple silicon.
+
+This is engineering evidence for ScopeProof's own installed workflow. It does not establish WCAG
+conformance, correctness, target-repository runtime behavior, authenticated reviewer authority,
+customer use, production use, accessibility by every user, or broad platform support. It does not
+advance Stage 1.
+
+The canonical Stage 1 counts remain:
+
+- 0/5 qualifying reviews;
+- 0/3 independent practitioners;
+- 0/3 public repositories;
+- 0/3 independently observed under-ten-minute completions;
+- 0/2 reuse-intent signals.
+
+## Implemented evidence slice
+
+- The installed-wheel Playwright regression now reaches the complete constructed primary path
+  with page-level Tab navigation, keyboard typing, Enter, and Space only. It no longer clicks the
+  demo disclosure or moves focus programmatically.
+- The regression requires the demo disclosure, demo-load button, reviewer input, criteria-confirm
+  button, and deterministic-analysis button to be enabled and intersect the viewport. It compares
+  unfocused and focused computed styles and requires a changed, nontransparent outline or shadow,
+  so always-on decoration and transparent focus styling cannot satisfy the assertion.
+- Existing 1280-by-720 and 390-by-844 contexts, isolated temporary home, loopback-only network
+  guard, browser-error assertions, evidence content, export controls, width checks, and process
+  cleanup remain in force.
+- CI now has a separate `compatibility-python-313` job. It does not rename the existing Python
+  3.11 or `verify` jobs. The new job runs the complete suite, both deterministic benchmarks, wheel
+  build/reinstallation, `pip check`, both CLI version commands, installed benchmarks, and exact
+  loopback workbench health on CPython 3.13. `verify` depends on the new job.
+- ScopeProof product and deterministic-gate source code did not change in this slice.
+
+## Keyboard-only installed-wheel evidence
+
+Command:
+
+```text
+uv run pytest -q -m browser tests/browser/test_packaged_workbench.py
+```
+
+Result on the final executable/test verification tree: `3 passed in 26.11s`. This includes the
+full installed-wheel browser journey plus pure regressions proving that unchanged always-on
+decoration is rejected and a changed visible outline is accepted. The complete suite on the same
+tree passed with `1,956 passed, 2 skipped` and 95.06% combined coverage. The hosted `verify` job
+also passed the packaged browser regression on this exact head.
+
+Earlier in the slice, the first new journey run failed because the helper required the Confirm
+button to be enabled before Tab blurred and committed the reviewer input. The DOM already
+contained the full reviewer value. Moving the enabled assertion to the point where keyboard focus
+actually reaches the target fixed that ordering defect. A subsequent independent review found
+that the original focus predicate could accept unchanged or transparent decoration; the final
+helper and the two pure regressions above close that false-pass path. No ScopeProof product change
+was required for either repair.
+
+The final regression repeated the full path in fresh 1280-by-720 and 390-by-844 contexts. It
+opened `Try ScopeProof`, loaded the deliberately constructed demo, entered the asserted reviewer
+label, confirmed normalized criteria, and ran deterministic analysis without pointer input or
+programmatic focus. Every primary target met the visible-focus and in-viewport assertions. The
+matrix retained missing-evidence guidance, the summary remained fail-closed at `Action required`,
+and Markdown, JSON, and CSV controls were visible and enabled. The final browser and external-
+request lists were empty.
+
+This proves the tested primary path. It does not prove every control or all possible tab orders.
+
+## Actual 200% Chrome zoom evidence
+
+- Browser: installed Google Chrome 151.0.7922.77.
+- Physical outer window: 3440 by 1328 pixels in both observations.
+- At reset 100%: device pixel ratio 1; CSS viewport 3440 by 1151.
+- At actual 200%: Chrome's toolbar exposed `Zoom: 200%`; device pixel ratio was 2; CSS viewport
+  was 1720 by 575; document horizontal scroll width was 1720.
+
+At 200%, a second exact-implementation-head journey used only keyboard navigation. Focus reached
+the demo disclosure, load button, reviewer input, confirmation button, and analysis button after
+8, 1, 13, 36, and 9 bounded Tab presses respectively. Each target was enabled, inside the current
+viewport, and showed either the three-pixel focus outline or the explicit focus shadow. The full
+evidence matrix, `5 · Summary & Export`, all three download controls, and `Review status: Action
+required` appeared. No named visible button, link, input, textarea, select, summary, or button-role
+element extended beyond the horizontal viewport. Chrome was restored to device pixel ratio 1 and
+the 3440-by-1151 CSS viewport after the run.
+
+This is real native-zoom evidence for the exact Chrome/macOS/display configuration above. It is
+not a claim about other browsers, displays, platforms, zoom levels, assistive technologies, or
+accessibility conformance. Responsive viewport resizing was not used as zoom evidence.
+
+## Python 3.13 evidence
+
+A genuine uv-managed `cpython-3.13.14-macos-aarch64-none` interpreter was installed from uv's
+managed Python distribution. On the implementation tree, a fresh wheel was built and installed
+with runtime dependencies into an isolated Python 3.13 virtual environment outside the checkout.
+
+The following checks passed:
+
+- `Python 3.13.14` interpreter identity;
+- wheel build and clean runtime-dependency installation;
+- `pip check` with no broken requirements;
+- installed package metadata and `scopeproof_core.__version__` equality at `0.2.3`;
+- `scopeproof --version` and `scopeproof-web --version` at `0.2.3`;
+- deterministic benchmark: 12 cases, 13 criteria, zero mismatches, zero must-have False Ready
+  outcomes, zero false blockers, and zero unexecuted declared categories;
+- comparison benchmark: 2 cases and zero mismatches; and
+- installed `scopeproof-web` exact loopback health response `ok`.
+
+The local result establishes Python 3.13 package, CLI, deterministic-gate, and health
+compatibility on this macOS host. The separate hosted job must pass on the PR before Linux-runner
+Python 3.13 evidence is claimed. Neither result is Linux desktop evidence.
+
+## Unsupported evidence rows
+
+| Row | Classification | Exact boundary |
+| --- | --- | --- |
+| VoiceOver or another real screen reader | Unsupported current evidence | macOS exposed the installed `com.apple.VoiceOver` service, but two bounded attempts to obtain a controllable, observable VoiceOver state timed out. No speech, caption-panel, focus-announcement, or reading-order evidence was available, so no screen-reader result is claimed. |
+| Windows desktop | Unavailable environment | No real Windows environment was available for build, installation, launch, or representative workflow completion. |
+| Linux desktop | Unavailable environment | No real Linux desktop environment was available. The headless Ubuntu CI job is package/runtime evidence only. |
+| Accessibility conformance | Unsupported claim | Keyboard, focus, native-zoom, DOM, and automated evidence cannot by themselves establish WCAG conformance or usability for every user. |
+
+## Verification status
+
+The focused browser regression, repository contracts, complete local suite, and local Python 3.13
+package/CLI/health checks passed on the final executable/test verification tree. Hosted CI passed
+the Python 3.11, Python 3.13, locked-environment, full verification, packaged-browser, and CodeQL
+checks on the same head. The opt-in informational ScopeProof evidence-review job skipped by design.
+The audit-only commit that records these results necessarily follows the tested head and does not
+change executable or test content.

docs/releases/v0.2.3-status-and-next-stages.md

@@ -162,10 +162,11 @@ remains zero independent use.
 | Missing GitHub patch integrity | Repaired on merged source | Absent, null, and empty patches make ingestion partial and identify the excluded path; malformed patch types fail with a bounded adapter error |
 | Reviewer-loop friction | Reduced on merged source | Reopened source identity, direct matrix navigation, insufficient-evidence rationale, same-head comparison, and comparison downloads have regression coverage |
 | R-003 holdout | Owner decision | Approve the outcome-blind cohort and criteria proposal before cohort generation; approve labels separately before scoring |
-| Keyboard and assistive technology | Unsupported current evidence | The installed flow completed with pointer automation, but the browser harness could not reliably demonstrate native keyboard activation; complete the packaged flow keyboard-only, with visible focus and a real screen reader |
-| 200% zoom | Unsupported current evidence | A 640 px responsive check showed no horizontal overflow, but the browser harness could not apply reliable page zoom; run the packaged workbench at actual 200% browser zoom without loss of content or operation |
+| Keyboard-only primary path | Supported local engineering evidence | The installed-wheel regression now completes the constructed path at 1280×720 and 390×844 with Tab, keyboard typing, Enter, and Space only; five primary targets must expose visible focus in the viewport. This does not prove every control or accessibility conformance. |
+| 200% zoom | Supported bounded local evidence | Installed Chrome 151.0.7922.77 reported native `Zoom: 200%`, device pixel ratio changed from 1 to 2, the exact-head keyboard path completed, and no named control was horizontally clipped on the tested macOS/display configuration. This is not responsive resizing or a broad browser/platform claim. |
+| VoiceOver or another real screen reader | Unsupported current evidence | The installed VoiceOver service was detectable, but bounded control attempts exposed no observable screen-reader state, speech, caption, focus announcement, or reading-order result. |
 | Python 3.11 | Remote CI evidence only | Python 3.11 CI passed at the exact merged PR head; no local Python 3.11 desktop flow was executed |
-| Python 3.13 | Unavailable environment | Run current-head gates on an actual Python 3.13 interpreter |
+| Python 3.13 | Supported local engineering evidence | A genuine CPython 3.13.14 environment passed clean wheel installation, `pip check`, both versioned CLIs, both deterministic benchmarks, and exact workbench health. The new hosted 3.13 job must pass before its Linux-runner result is claimed. |
 | Windows and Linux desktop | Unavailable environment | Build, install, launch, and complete representative workflows on those operating systems |
 | Stage 1 | External evidence gate | A non-owner supplies a real public PR, public requirements, source-owner confirmation, exact head SHA, saved review, and genuine outcome |
 | Stage 2 | Stage gate | Every Stage 1 target passes before repeat-use and commercial-discovery evidence is collected |
@@ -191,6 +192,11 @@ lock check, both deterministic benchmarks, clean installation, installed
 benchmarks, workbench health, browser journey, and responsive-site checks.
 Engineering work does not advance Stage 1.
 
+The current [accessibility and platform evidence audit](../audits/accessibility-platform-evidence/verification.md)
+adds a durable installed-wheel keyboard/focus regression, bounded native Chrome 200% evidence, and
+genuine local Python 3.13 package/CLI/health evidence. VoiceOver and real Windows/Linux desktop
+workflows remain unsupported. These engineering checks do not advance Stage 1.
+
 ### Stage 1 — genuine public alpha
 
 Status: `waiting_for_inbound_public_alpha_submission`.

docs/superpowers/plans/2026-08-09-accessibility-platform-evidence.md

@@ -0,0 +1,81 @@
+# Accessibility and platform evidence implementation plan
+
+> Execute continuously on `codex/accessibility-platform-evidence`. Do not merge, release, tag,
+> begin outreach, generate R-003, retune R-002, or claim Stage 1 progress.
+
+**Goal:** Add durable installed-wheel keyboard/focus evidence and genuine Python 3.13 compatibility
+evidence while recording native zoom and unsupported environments without overclaiming.
+
+**Architecture:** Keep product and gate code unchanged unless a reproduced defect requires a
+test-first repair. Extend the packaged Playwright boundary for keyboard interaction, add an
+isolated Python 3.13 CI compatibility job, and update evidence documents only from observed runs.
+
+---
+
+## Task 1: Establish the bounded environment evidence
+
+- Verify branch, base, status, GitHub authentication, checks, available browsers, Python
+  interpreters, VoiceOver availability, and absent Windows/Linux desktop environments.
+- Run the constructed primary path in real Chrome at native 200% zoom using keyboard input.
+- Require Chrome's visible 200% indicator plus matching pixel-ratio/viewport metrics, inspect
+  horizontal control clipping, and restore 100% zoom.
+- Attempt VoiceOver only through a bounded observable control path. If no trustworthy state is
+  exposed, record the exact limitation and stop that probe.
+
+## Task 2: Replace partial keyboard evidence with a durable regression
+
+**File:** `tests/browser/test_packaged_workbench.py`
+
+- Add a helper that reaches a locator with page-level Tab presses only.
+- Assert the active target is enabled, intersects the viewport, and has visible focus styling.
+- Open the demo disclosure and activate Load, Confirm, and Run using keyboard input only.
+- Enter the reviewer field through the keyboard path.
+- Preserve the existing installed-wheel, isolated-home, network, error, viewport, content, export,
+  and teardown assertions.
+- Run the focused packaged-browser test. If it exposes a product defect, preserve the failing test,
+  apply the narrowest repair, and rerun it.
+
+## Task 3: Add genuine Python 3.13 compatibility coverage
+
+**File:** `.github/workflows/ci.yml`
+
+- Add a separate `compatibility-python-313` job without renaming existing protected jobs.
+- Install on CPython 3.13, run the complete suite and both deterministic benchmarks, build and
+  reinstall the wheel, check imports and both versioned entry points, rerun both installed
+  benchmarks, launch the installed workbench on loopback, require exact health `ok`, and clean up.
+- Make `verify` depend on the new compatibility job.
+- Obtain a genuine managed local CPython 3.13 interpreter and repeat the package, CLI, health, and
+  deterministic-gate checks locally.
+
+## Task 4: Update exact evidence documents
+
+**Files:**
+
+- `docs/audits/accessibility-platform-evidence/verification.md`
+- `docs/releases/v0.2.3-status-and-next-stages.md`
+
+- Record the exact tested host, Chrome zoom signals, keyboard path, focus assertions, Python 3.13
+  interpreter, commands, and results.
+- Classify VoiceOver as unsupported unless observable screen-reader evidence was obtained.
+- Keep Windows and Linux desktop unsupported; distinguish the Linux headless CI job.
+- Preserve every trust boundary and the exact Stage 1 zero counts.
+
+## Task 5: Run full verification
+
+- Run Ruff.
+- Run the complete suite with combined coverage at or above 95%.
+- Run repository contracts and both deterministic benchmarks.
+- Build the wheel twice and require matching SHA-256 values.
+- Install the final wheel into fresh Python 3.12 and 3.13 environments; run `pip check`, package
+  imports, both version commands, both benchmarks, and exact loopback health.
+- Run the installed-wheel browser regression with the pinned Playwright driver.
+- Check distribution inventory, `git diff --check`, status, and preservation of `.coverage 2`.
+
+## Task 6: Publish for owner review
+
+- Stage named files only and commit intentionally.
+- Push the feature branch and open a ready-for-review PR against `main`.
+- Monitor CI, CodeQL, and review feedback to completion.
+- Diagnose failures systematically; fix only confirmed in-scope defects with regression coverage.
+- Recheck the final diff, commits, mergeability, exact head, check conclusions, unsupported
+  environments, and owner decision required.