# Manual device and accessibility acceptance

Automated tests do not establish screen-reader announcements, native clipboard interoperability, physical touch behavior, presented frame rate or true peak memory. Use this protocol on the exact source revision being evaluated, with public sample data. Record failures and evidence before expanding support claims. No result is implied by this document.

## Local command reference budgets

Use two initial workload profiles: 10,000 allocated rows for frequent local commands (50 ms p95 command/timer-delay target), and 100,000 rows for explicit bulk commands (250 ms target). These are engineering targets for evaluation, not supported-device guarantees or a production SLA. The paste profile changes one editable numeric column across all rows; sort uses two keys and filter tests equality. A host with more columns, expensive parsers/policies, groups or application subscribers needs its own profile.

After building the engine, run the opt-in warmed measurement on a quiet machine:

```powershell
$env:ACHERON_RESPONSIVENESS='1'
npx playwright test --config benchmark.config.mjs tests/benchmark-responsiveness.mjs --workers=1 --reporter=json
```

On POSIX shells prefix the command with `ACHERON_RESPONSIVENESS=1`. Each profile performs two warmup trials and 30 measured trials. The attachment retains raw command/timer durations, median and nearest-rank p95, correctness checks and informational budget results. Default CI skips this opt-in timing workload; other correctness gates remain active. A budget miss is reported rather than failing correctness. Establish repeated stable-host baselines before enforcing a timing gate.

Keep small cell edits on the ordinary synchronous API. Treat 100,000-row sort/paste/history as explicit bulk actions: show a pending state before starting (allow the browser a paint opportunity), prevent duplicate submissions and preserve atomic validation/history. A pending state cannot make a synchronous command interruptible. If the host needs interaction while sorting large datasets, use server-owned sorting/filtering through the existing read-only async source query lifecycle. Do not slice one atomic paste into separate commands without deliberately accepting different rollback/history behavior. The Unreleased `createRemoteDataSource` supports bounded optimistic writes against a full snapshot; it is not a paged write adapter. Validate commit retries, conflict resolution and resync separately using the [editing and remote guide](editing-and-remote.md).

## Environment record

Copy this record for each combination:

```text
Source revision:
Package versions/source preview:
Date:
Tester:
OS and version:
Browser and version:
Assistive technology and version (or none):
Device, CPU/RAM, display refresh rate and DPR:
Dataset shape, column types, groups/merges/frozen counts:
Result per case: PASS / FAIL / NOT RUN
Observed and expected behavior:
Reproduction and evidence location:
```

Start with NVDA + Firefox/Chromium on Windows, VoiceOver + Safari on macOS, and physical iOS/Android touch devices. These combinations are planned coverage, not verified support.

## Keyboard and screen readers

1. Focus the grid. Navigate with arrows, Home/End and Ctrl/Meta+Home/End. Confirm row/column, value, selection and read-only announcements match visible data.
2. Use Shift to extend ranges, select whole rows/columns and navigate frozen/grouped/projected rows. Repeat after sorting, filtering and collapsing groups; announcements must match stable record identities.
3. Open an editor with F2. Submit an invalid draft, correct it, save with Enter and cancel a second edit with Escape. Confirm the field name/error is announced, focus returns correctly and cancel preserves values.
4. Open the menu with Shift+F10. Search and navigate options, including disabled options. Apply and cancel choices; check focus after Escape/Tab.
5. Repeat with reduced motion. Destroy/remount the grid; removed controls must not remain focusable or produce duplicate announcements.

## Native clipboard

1. Copy/paste a rectangular range between two grid instances and a native text/spreadsheet application. Include Unicode, quoted tabs, multiline values and empty cells; verify exact resulting values.
2. Test same-grid cut/move, copy after cut, Escape cancellation, permission rejection and undo/redo. Rejected paste must preserve source and destination values/history.
3. Paste while a text editor is active; its native text selection/clipboard must remain usable. Check permission-denied and unavailable clipboard behavior through the actual browser UI.
4. Repeat with public rich-text/HTML input and supported formats; confirm unsafe content stays inert and unsupported formats have the documented fallback.

## Physical touch and Safari

1. Scroll the page and grid in both axes. Select a cell/range, drag a header where enabled, and edit without accidentally trapping page scrolling.
2. Exercise frozen panes, narrow layouts, on-screen keyboard, rotation, menus and choice editors. Confirm targets remain reachable and the editor stays aligned with its cell.
3. Background/resume and destroy/remount; confirm stale touch gestures, menus and listeners do not remain active.

## Responsiveness and memory

1. Select representative host workloads and an acceptable interaction-latency budget **before** comparing timings. Record allocated rows/cells, edit shapes, sort/filter criteria, cache limits and update rates.
2. Warm up the application, then collect at least 30 samples per workload on a stable runner. Preserve raw durations, operation errors, median/p95 and frame evidence; do not discard slow samples without a recorded reason.
3. Capture a browser performance trace through scroll, sort, paste, undo, live bursts and remote query changes. Report presented frames only when the tool actually observes them; requestAnimationFrame cadence alone is insufficient.
4. Sample browser/renderer process memory and heap during allocation, large commands and teardown. Record sampling rate and process identities. Report a **maximum observed sample**, not true transient peak. Use an OS/native profiler if a true peak or GPU allocation claim is required.
5. Repeat mount/destroy and query/cache churn. Compare DOM/listener/cache counters and retained heap after warmup; document host-owned subscriptions and object URLs separately.

## Release decision

Any failure needs the exact revision, reproduction, expected/observed result and evidence. Keep NOT RUN items open. Timing gates require a stable runner, a host-approved budget and repeated baseline data; the existing CI gates enforce correctness and resource bounds without inventing a hardware-independent latency threshold. Package publication and production deployment remain separate actions.

## Source editing and remote follow-up

At the recorded source revision, verify special paste in a read-only-formatting cell, transpose/skip-empty with protected destinations, and atomic rollback/history. Hide edge/interior/all axes, recover via Shift+F10, restore saved state and confirm retained values/sizes. Check localized search/menu/editor/validation announcements in English, Vietnamese and a host pack, with numeric ARIA/display matching the locale while editor values remain raw. Run desktop and narrow viewport checks.

Against the actual application server, simulate a lost accepted response and retry the same mutation ID, concurrent revision conflict, rejected validation, disconnect during load/commit and resync after server-side reorder. Verify authorization, durable idempotency, explicit draft review and identity capture/engine refresh. Deterministic transport fixtures alone do not close this server acceptance gate. The [source guide](editing-and-remote.md) defines the lifecycle.