fix(note-arrays): size send paths to real dimensions + harden tests [gate for #1185]

[jeffredodd]

Why this PR

This is the human-in-the-loop gate for #1185 (Note Arrays). It targets next, so merging it updates #1185. Goal: when this merges, #1185 is actually 100% correct for arrays + multi-board, with tests you can trust.

I rebuilt the next (04c4f08b) dev stack from scratch (with the new mock Cloud board on :19200) and verified the feature at runtime — not just by reading tests. That surfaced a ship-blocker the existing tests hid.

🔴 Blocker fixed: note-array boards 500 on every send path

get_dimensions() only knows flagship/note and raises ValueError on note_array. Five send paths called it. Reproduced live:

POST /templates/render/live  (note-array board)  ->  HTTP 500
  File "src/api_server.py", line 7419, in render_template_live
    dims = get_dimensions(device_type)
  ValueError: Unknown device type: note_array. ... use resolve_dimensions().

A user cannot send to a note-array board via the live editor, /pages/{id}/send, /settings/active-page, or /displays/{type}/send. The send-endpoint tests patch("…get_dimensions") and hardcode rows=6, cols=22, so they could never catch it.

Fix: all five send sites use resolve_dimensions(device_type, notes_wide, notes_tall). After the fix the same live send returns 200 and lands a 3×60 grid on the mock.

What's in this PR (all verified green)

Change	File
5 send paths size to real note-array dims (was 500)	src/api_server.py
Guard from_dict against a JSON-null note_array_token (None.strip() crash)	src/devices.py
New regression test: live send to note-array board, resolver NOT mocked, asserts 3×60 grid	tests/test_api_extended.py
Repoint 7 send tests get_dimensions→resolve_dimensions	tests/test_api_coverage.py
Make 2 note-array Cloud-URL tests hermetic (pass in docker exec … pytest, not only CI)	tests/test_board_client.py
Run the integration-tests/ mock-cloud contract suite in CI (was never collected)	.github/workflows/ci.yml
i18n: localize size-indicator preset labels (English leaked into visible text and aria-label in 13 locales)	web/src/components/board-size-indicator.tsx
Type contract: add schedule_enabled + port to BoardInstance (server emits them; TS type was incomplete)	web/src/lib/api.ts

Results: Python 4102 passed / 11 skipped (was 4099 + 2 failing), integration 25 passed, web vitest 1166 passed, prettier/eslint clean.

Verification performed against the live next build

• Cloud transport (14/14, real BoardClient → real mock): URL + X-Vestaboard-Token, body exactly {"characters": grid} with transitions stripped, 3×60 lands, read round-trips, 15s throttle blocks then allows, per-board token isolation holds.
• Auto-detect (13/13): every preset + custom 8×8 + flagship + note classify correctly; 404/422 paths; token masked.
• Web E2E (78 passed): note-arrays regression, note-pages, multi-board, multi-board-schedule (incl. board_id filtering + board-switch isolation), setup wizard, device-mismatch.
• Pages schema migration v3→v4: correct, idempotent, no data loss.
• TS↔Python contract (full sweep): in sync except the two BoardInstance fields fixed here.
• i18n parity: all 14 locales carry every note-array key (no missing/orphan); the one real leak is fixed here.

✅ Investigated and cleared (were flagged, verified NOT bugs)

• Settings migration "drops schedule.enabled": not a bug. is_schedule_enabled() falls back to the global mirror when there are no boards, and the migration stamps per-board schedule_enabled when boards exist (and never deletes the global value). No upgrade data loss.
• "Re-POSTing boards drops schedule_enabled/port": not data loss. The Settings UI spreads the received board object ({...b, ...updates}), so runtime fields survive even though the TS type omitted them (now added for honesty).

⚠️ Still your call (out of scope / not arrays-correctness)

1. Multi-board driving is primary-only (confirm intended). The polling loop drives only the first board (src/main.py:733) and the dashboard preview ignores the sidebar selection (active-page-display.tsx); also PUT /settings/active-page is primary-board-scoped (no board_id). Looks like later per-board-epic scope, not claimed-broken here.
2. Unpinned ruff (ruff>=0.9.0) → nondeterministic CI format-check. ruff 0.15 reformats 64 src/ files of pre-existing code. Can fail ruff format --check for any new PR to next. Recommend pinning ruff + a one-time reformat (separate PR).
3. detect-size "unclassifiable" 422 branch is dead — the read layer pre-filters odd grids to None, so users get a misleading "blank or unreachable" message for an unsupported size.
4. Lower-priority test-quality gaps (finders, not all verified): scaled-board-display fit-mode scaling untested; seam gap not asserted; page-grid-selector renders note-array thumbnails at 1×1; plugin note-array cache tests assert only fetch_count.

🤖 Generated with Claude Code

[jeffredodd]

Multi-board scope clarification (resolves flagged item #1): the per-board driving limitations I noted are tracked as separate open epic issues, not gaps in #1185's claims:

• Polling loop drives primary board only (src/main.py:733) → Per-board #2: per-board display engine (DisplayService → runtimes) #1243 (per-board display engine) + Per-board #3: per-board send routing (endpoints + MQTT + triggers) #1244 (per-board send routing)
• Dashboard preview / manual active-page ignore the sidebar-selected board (active-page-display.tsx; PUT /settings/active-page has no board_id) → Per-board #6: dashboard per-board active display #1247 (dashboard per-board active display)

#1185 ships the multi-board foundation (per-board settings + schema migration, per-board schedule storage/filtering by board_id, sidebar selector + CurrentBoardProvider, unified Schedule page). Full per-board driving is intentionally deferred to #1243/#1244/#1247. The single product decision for this release: ship the foundation now with driving to follow, or hold #1185 until those land.

[github-actions]

No a11y concerns — this PR is a net improvement.

The only web UI change is in web/src/components/board-size-indicator.tsx. It replaces the hardcoded English preset labels that were previously leaking from NOTE_ARRAY_PRESETS[].label into the ariaLabelWithLayout aria-label string with properly localized strings via tSettings(presets.${presetId}).

Verified:

• Preset IDs align with i18n keys. The five IDs in NOTE_ARRAY_PRESETS (2_wide, 4_wide, 2_tall, 4_tall, 2x2_grid) match exactly the keys under displaySettings.presets in the PR head's en.json. No missing or misspelled keys.
• boardSizeIndicator keys are intact. ariaLabel, ariaLabelWithLayout, and custom all exist in en.json. The ariaLabelWithLayout string ("{rows} rows by {cols} columns, {layout}") produces a coherent screen-reader announcement once layout is filled with the localized preset name.
• No hardcoded English strings introduced. Both the visible preset label and the aria-label text are now derived from next-intl.
• All 14 locale files already carry displaySettings.presets — no locale drift introduced.
• Semantics unchanged. The role="img" + aria-label pattern on the indicator wrapper was correct before and remains correct; the decorative · separator sits inside the img role and is hidden from AT via the aria-label override.

This is a straightforward i18n fix that also improves the screen-reader experience for non-English users.