XStats/ARCHITECTURE.md

# XStats architecture

macOS 14+ menu-bar system monitor. Swift 6 language mode, AppKit status item and panels
hosting SwiftUI. The Xcode project is generated from `project.yml` by XcodeGen;
the app's testable components live in the local Swift package. For directory layout, build, test,
and release instructions, see [DEVELOPMENT.md](DEVELOPMENT.md).

## Sampling

`MetricsHub` is an actor that runs one loop. `AppModel.demand` describes what is on screen —
which menu-bar items are enabled, whether the panel is open and on which tab — and the hub
samples only that: CPU always; memory and network cheaply; GPU, disk, processes, sensors and
fans only when a visible surface needs them. Disk is read at most every 30 s, battery every
10 s. The loop pauses on screen sleep, system sleep and session switch.

The sandboxed WidgetKit extension samples CPU, memory, disk, and battery through `Metrics`
independently, so it remains useful when the app is not running.

Things that are easy to get wrong and are handled on purpose:

- `host_processor_info` returns kernel-allocated memory; it is `vm_deallocate`d every sample.
- `getifaddrs`' `if_data` counters are 32-bit and wrap at 4 GiB; network uses
  `NET_RT_IFLIST2` and `if_data64`.
- Memory follows Activity Monitor: app memory is `internal − purgeable` pages; the page size
  comes from `host_page_size` (16 KB on Apple Silicon).
- Process CPU time and wall time are both in mach absolute units, so their ratio needs no
  timebase conversion.
- Core types come from `hw.perflevelN`; logical CPUs are numbered from the lowest
  performance level up.

## SMC

Temperatures are not hard-coded per chip. On first use the sampler enumerates every SMC key
once (≈3,800 keys in about 10 ms on an M5 Max), groups `Tp*`/`Te*` as CPU, `Tg*` as GPU,
`Tm*` as memory, `TB*` as battery and `Ts0P`/`Ts1P` as palm rest, keeps keys whose first
reading is plausible, and samples at most 12 per group.

Fans are read without privileges. Writing needs root and goes through the helper. On M5
the mode key (`F0md`) accepts a direct write; M1–M4 first need `Ftst=1` and a pause while
`thermalmonitord` lets go. The firmware records only manual or automatic, not who set it, so
a fan in manual mode that XStats did not set is shown as controlled by another program.

## Helper

Protocol 5 removes the identifier-only ad-hoc authentication fallback. Ad-hoc apps refuse helper
registration and unregister any previously registered helper without connecting to it. Fan and
lid-closed controls require team signing; maintenance operations keep their existing admin-prompt fallback.
Readiness requires an authenticated version handshake, not just an enabled SMAppService registration.
For older helpers, the client unregisters the service before re-registering the current bundle and
verifying its version. Unregister/handshake failures leave privileged calls disabled and surface an error.

Registered with `SMAppService.daemon`; `RunAtLoad` so that an unclean exit is repaired at
boot. The XPC interface is a fixed list of operations — no arbitrary commands — and each
connection gets `setCodeSigningRequirement`, derived from the helper's own signing team rather
than a hard-coded team ID. State that must be undone (manual fans, disabled sleep) is persisted
to `/Library/Application Support/XStats/helper-state.plist` and
reverted when the last client disconnects or at the next start. The helper exits after 30 s
without clients.

## Cleanup

Every rule lists candidate items; every item passes `SafetyGuard` twice — at scan and again
right before deletion, because apps start in between.

- Deny by default: an item must sit strictly inside an allow-listed root under the home
  folder, never be the root itself, contain no `..` or control characters, and still pass
  after symlinks are resolved (resolution can only reject, never allow).
- Under `Application Support`, only folders named like caches (`Code Cache`, `GPUCache`,
  `CacheStorage`, …) are allowed.
- Any path component matching a protected keyword (keychains, password managers, VPNs,
  cookies, history, …) is rejected.
- Reverse-DNS cache folders whose app is running are skipped; browser rules are blocked while
  the browser runs; items modified in the last two minutes count as in use.

Regenerable caches are deleted outright; user files go to the Trash. Each action is appended
to `~/Library/Logs/XStats/cleanup.log` as a JSON line.

## Disk tools

The disk page is also where users act on the disk. `DiskToolsController` (`XStatsUI/State`) fronts four
pieces of read-mostly logic in `Cleaner/DiskTools.swift`, each testable without the UI:

- `SpaceScanner` walks a root (the home folder) once, sums every top-level entry and keeps the largest files;
  packages (`.app`, `.photoslibrary`, …) count as one item. It never follows symlinks and skips folders it
  cannot read. `canTrash` allows Trash only for files inside the home folder, outside `Library`, with no hidden
  path component, checked again after resolving symlinks; everything else only gets "reveal in Finder".
- `VolumeVerifier` runs `diskutil verifyVolume /` (read-only, no privileges) and reduces the output to OK /
  problem plus the offending line.
- `LocalSnapshots` parses `tmutil listlocalsnapshots /`. Deleting goes through the helper
  (`deleteLocalSnapshots`, protocol version 4) and falls back to a one-off administrator prompt; both paths only
  accept identifiers shaped like `2026-09-14-120000` (`HelperShared/LocalSnapshotCommand`).
- `MountedVolumes` lists browsable non-root volumes; ejecting uses `NSWorkspace`.

## Menu bar, popovers and main window

`MenuBarController` owns the status items. In the *separate* layout every enabled metric gets
its own `NSStatusItem` (created in reverse so they read left to right) and opens a 320 pt
popover for that metric; in the *combined* layout a single item opens `CombinedPopoverView`: a status
overview with one row per enabled metric, plus tabs that switch to each metric's full popover content
(`PopoverDetail`, shared with the separate layout). While a detail tab is showing, `AppModel.openPopover`
is set to that metric so sampling matches the standalone popover.

Popovers are borderless, non-activating `NSPanel`s. The SwiftUI tree is created on open and
destroyed on close, so a hidden popover costs nothing. Height comes from measuring a flat,
scroll-free copy of the same view; placeholders keep that height stable until data arrives, and
the window is resized without animation because animating it makes SwiftUI re-lay out every
frame. Charts are drawn with `Canvas`, not Swift Charts.

The main window reuses the popover content with `isDetailPage` set (all sections, taller charts)
next to the dashboard and tool pages. Windows use a transparent, full-size-content title bar with
an empty compact toolbar, so the traffic lights sit on the same ground colour as the sidebar and
line up with the 40 pt page header; the app switches to a regular activation policy while a window
is open and back to accessory when all are closed.

## AI process explanations

`ProcessExplainer` uses only the on-device Apple Intelligence model. It checks framework and
macOS support before exposing the action, then checks model availability at request time. If
the device, system setting or model is unavailable, the explanation card shows the reason and
does not call another provider. Successful requests stream local output and can be cancelled.
The model receives process identity, path, signer and resource usage; nothing is sent to a CLI
or network service. Model response instructions follow the resolved application language.
## Calendar

`CalendarMenuBarController` owns a separate, opt-in status item, independent of the metrics
layout and `MetricsHub`. It reuses `StatusPanel`, updates the displayed date every minute and
on day/time-zone changes or wake, and stops its timer when disabled. `CalendarPopover` uses a
six-week Gregorian grid (1900–2100) and month/year navigation. Clicking a day opens an almanac
page in the same panel; returning preserves the selected date and month. `CalendarAlmanac`
loads Tyme's daily advice, deities and hour fortunes only for the selected day and caches the
value model by civil-date ID. The twelve-hour grid uses early Zi through Hai, excluding the
next day's late Zi entry at 23:00; each cell exposes its exact time range in a tooltip. Browsing
history preserves the selection across midnight; a selection on today follows the new day.

`CalendarEngine` is the sole adapter for the pinned MIT-licensed Tyme4Swift 1.5.0 dependency.
All Tyme calls run on MainActor because upstream exposes mutable static tables; only value
models leave the adapter. Civil dates use the system time zone and calendar day arithmetic,
not 24-hour intervals. Tibetan conversion is guarded to the upstream Gregorian coverage
1951-01-08 through 2051-02-11. Holiday data currently ends in 2026; unknown years show a notice.
Plum-rain dates describe traditional calendar rules, not observed or forecast weather.

Calendar enablement, visible features and week start persist through `AppSettings` and the
optional fields in `SettingsDocument`. Older backups preserve those settings. Additional
calendars default off and appear only in selected-day details. There is no event access or
network request at runtime. Use `--snapshot <directory> --calendar-only` for deterministic
light/dark calendar screenshots without starting unrelated samplers or scans.

## AI usage and quotas

The local Codex provider reads JSONL rollouts from CODEX_HOME (default ~/.codex),
including sessions and archived_sessions. It never reads credentials or calls quota APIs;
independent quota providers do that only while the master AI Usage switch is enabled.
Parsing runs in a background actor pinned to a dedicated serial executor, so a cold scan (tens of
seconds over a multi-gigabyte log directory) never occupies a Swift concurrency cooperative thread
and cannot stall per-second metric sampling. It uses bounded line buffers and a SQLite checkpoint store at
~/Library/Application Support/XStats/ai-usage.sqlite. Each source/path row atomically stores file
identity, size, mtime, newline-aligned byte offset and Codable parser/aggregate state. The preview
state is stored only when the final line is still unterminated, so the common case keeps one copy.
Rows whose files were not seen in a completed scan are pruned, which covers rollouts moved to
archived_sessions, deleted project directories and superseded parser namespaces. Unchanged
files reuse stored results even after restart. Appends validate prefix/boundary hashes and read
only the suffix; truncation, replacement, parser namespace or time-zone changes rebuild that file.
An unterminated final line is previewed separately from committed state, so completing it cannot
double-count usage. No raw log lines, prompts or credentials are persisted. Directory metadata is
still enumerated on each scheduled refresh; only currently discovered files contribute to totals.
The dashboard defaults to the last 365 local-calendar days by model, and shows input/output tokens, cache hit
rate, record count, daily totals and model ranking. The menu bar shows today's token total.
Cached input is a subset of input; reasoning is a subset of output, so neither is added twice.
Rows are aggregated by day and model before leaving the provider, and ordered by day then model, so
`ModelTokenUsage.id` stays unique and an unchanged corpus produces an equal report.
Duplicate session IDs and unchanged cumulative snapshots are excluded. Child replayed history
only seeds the cumulative baseline until the first `task_started`, which always follows that
history; `started_at` is not compared against the session creation time, because a subagent's
`started_at` is its parent turn's start and recent Codex builds omit the field entirely.
Unknown models remain unknown.
Local logs cannot reliably identify the paying account or usage on other devices. No cost,
subscription limits or HTTP success rates are inferred from them. The separate quota
snapshot instead queries Codex `wham/usage` and Claude `api/oauth/usage` using CLI OAuth tokens.
It maps five-hour and seven-day windows, plus Claude's model-specific weekly windows, without
combining them with local token totals. Credential discovery is repeated on each refresh; Codex
reads `auth.json` under CODEX_HOME or the standard locations, while Claude reads its CLI credential
file or a non-interactive Keychain item. Tokens are not persisted by XStats. Each provider's last
successful quota snapshot is stored in a separate table of `ai-usage.sqlite`, without credentials.
The controller restores it before the first network request and labels it with its fetch time.
Requests use isolated URL sessions and reject redirects so bearer tokens cannot be forwarded.
Missing or expired credentials clear that provider's quota display; transient errors keep the last
successful result marked stale across restarts. Authentication failures or removal of a manual
account delete that provider's cached snapshot. Both quota readers stop when AI Usage is disabled;
disabled providers are hidden while their cached snapshot remains available on re-enable.
Codex and Claude Code each have an optional, independently configured Sub2API source, tried only
when that provider's direct quota fetch fails. Each HTTPS base URL, admin email and account ID remain
in local preferences outside settings backups; passwords use separate Keychain items. The fallback
signs in to its configured server, reads the account record to verify its platform (`openai` or
`anthropic`), then reads `/api/v1/admin/accounts/{id}/usage` with `source=active&force=false`.
If the detailed endpoint is unavailable, the account's `extra` fields provide coarser five-hour and
weekly windows. Claude may also expose Sonnet and Fable weekly windows. Credentials and quota results
are never sent to XStats servers; the UI marks fallback results as Sub2API.
When one provider has a subscription snapshot, the menu bar AI reading shows its name and remaining
percentage. It prefers the weekly window (or five-hour window if weekly is unavailable); the tooltip
shows the chosen window and reset time without crowding the 22-point menu bar.
When both sources are enabled and either has quota data, the item shows a compact named reading
for each source; one without quota data shows a dash. The tooltip lists reset times and missing
data by source. Without any subscription snapshot it keeps the local daily Token reading.
The AI menu bar item follows the global or per-item style. Ring, pie, meter, and dot styles
show remaining quota beside its percentage for each available source; a source without quota
shows a dash rather than a zero gauge. Since quota history is not stored, global history or
line styles use a current-value ring for this item.
Icon style uses the bundled OpenAI and Claude logo shapes for the corresponding quota source,
including a dash when that source has no quota; aggregated local-only tokens keep the generic AI icon.
The quota detail card presents remaining allowance (not used allowance) and, when the server
provides a reset date, a separate bar for elapsed time in that five-hour or seven-day window.
Allowance turns red as it runs low; the reset bar turns green as reset approaches and updates
while the view is open.
The UI follows CC Switch's filter/summary/trend/model-table organization, implemented in SwiftUI.
The desktop toolbar groups source controls and actions. In the compact menu-bar popover, its header
owns refresh and AI Usage settings; the content shows a source switch only when multiple sources
are enabled. Token totals use compact notation with
exact hover/VoiceOver values. Cache details are
disclosed on demand. AI usage settings open in a separate window so users can copy account details
from another app without closing the form. Unsaved Sub2API fields stay in memory for the current
app session across page changes; passwords are not written to preferences. Since source data is aggregated
by day, all three activity modes cover the same trailing 365 days: daily uses one cell per day;
weekly uses a seven-cell-high bar per calendar week; cumulative sums usage from the window start
through the selected week and uses the same bar grid. Zero weeks have no lit cells. Partial weeks
include only days inside the window through today. The mode switch lives in the Token activity
header, not in a date-range toolbar. Summary and model ranking follow the mode: daily means today,
weekly means the current calendar week, and cumulative means the current calendar month (all through
now, using the local calendar). The heatmap independently keeps its 365-day window. Pointer hover
immediately shows exact daily/weekly/cumulative counts above the chart; the same label is available
to accessibility. The narrow popover lays cells out at their design size and scrolls the year grid
horizontally, initially at the recent end; the wide panel shrinks cells to fit instead.
Numbers and dates follow the app language rather than the system locale.
Pointer selection fades only the selected control background (180 ms); keyboard selection and
background data refresh do not animate the data layout. Reduced motion uses the existing quick fade.

Claude Code is a second local provider. It scans projects/ recursively (including subagents) under
CLAUDE_CONFIG_DIR when set, otherwise ~/.claude and XDG_CONFIG_HOME/claude (default ~/.config/claude).
A checkpoint keeps every assistant event for its file, because the retention window is recomputed from
the current date on each fetch: a stored cutoff would freeze at the first scan and keep aged-out events
forever instead of dropping them.
Assistant message IDs are deduplicated across files: completed messages win over partial snapshots,
then the greater output count wins. Claude input is normalized as fresh input + cache reads + cache
creation; nested 5-minute/1-hour creation counts are a fallback, never added to the aggregate twice.
Independent local source switches enable Codex and Claude Code (both default on under the master
opt-in). Disabled sources are not scanned or included in cached reports/menu totals, and are hidden
from the source selector. Disabling the selected source resets the filter to all enabled sources.
Both switches may be off, in which case polling stops and the page offers the source settings.
The menu bar totals enabled sources for today; enabling the menu-bar item also turns scanning on,
since that item is the most natural place to discover the feature.
Refreshes are queued rather than dropped: a refresh arriving while another is in flight runs after it,
so rebuilding the polling cadence cannot mistake a skipped call for a completed initial refresh and
then sleep out the whole interval. Stopping cancels the in-flight scan. These settings are local and not synced through WebDAV.
Cache creation is shown separately but already included in the input total. Claude credentials,
account data, Cowork containers and subscription limits remain outside the local scanner's scope;
the quota reader has no access to session-log content.

## Network details

`NetworkController` runs only while it is needed:

- **Connection probe** — an unprivileged `SOCK_DGRAM` ICMP echo (`ConnectivityProbe`), matched on
  sequence number and a random payload token because the kernel rewrites the identifier.
- **Interface and addresses** — `SCDynamicStore` / `SCPreferences` for the primary and physical
  service, `getifaddrs` for addresses, CoreWLAN for signal and rate.
- **Public IP** — Cloudflare trace (ipify as fallback) for the address only. Location, ASN, network
  type, native / broadcast and the cleanliness score come from `cleanip.io/cli?json=1`, queried directly
  from each Mac by `PublicAddressLookup`. The endpoint only reports the caller's own address, so
  `AddressFamilyRequest` opens one Network.framework connection pinned to IPv4 and one pinned to IPv6
  (URLSession cannot choose the family) and speaks plain HTTP/1.1 over TLS; when a pinned connection
  cannot be set up it falls back to URLSession and keeps the answer only if it is for the same family.
  If cleanip.io sees a different exit than Cloudflare (split-routing proxies), its address is shown.
  Results are cached per IP for an hour in memory and for a week on disk (failed lookups are not
  cached), and a 429 stops further calls until the next UTC day. Country
  codes are validated before being used as flag file names. `server/geoip/` still holds the systemd
  timer that syncs MaxMind GeoLite2 onto `getopenstats.com/geoip/` (account and key in
  `/etc/openstats/maxmind.env` on the server); the app no longer downloads those files, they are kept
  for other uses.
- **Per-process traffic** — cumulative bytes from `/usr/bin/nettop`, diffed between samples.
- **DNS** — `networksetup -setdnsservers` through the helper (protocol 3), which re-validates the
  service name and every address; without the helper, a one-off administrator prompt runs the
  same fixed command.

## Battery and Bluetooth

The battery item reuses `BatterySampler` (IOKit power sources, sampled every 10 s while the item is in the
menu bar or its popover / page is open). `BatteryPopover` is both the popover and the main-window page
(`DetailPage`). The 24-hour charge curve comes from the history database, which gained a `battery` column
(migrated with `ALTER TABLE` on first open); `HistoryRecorder.loadBattery()` queries it independently of
the History page's range. Macs without a battery show only Bluetooth devices, and the menu-bar segment
falls back to the Bluetooth device with the lowest battery.

`BluetoothController` (`XStatsUI/State`) wraps `BluetoothBatteryReader`, which shells out to
`system_profiler` and takes a second or two, so it polls only on demand: every minute while the battery
popover / page or the System page is open, every five minutes when the menu bar needs it (the
low-battery hint or a Mac without a battery), otherwise not at all. `AppModel.bluetoothDemand` derives
that from the same visibility state as `demand`.

## Online updates

`UpdateController` posts the current version and a hashed random installation ID at launch and daily.
China-region locales prefer `https://x-stats.china.12306.work/api/v1/update/check`; other locales prefer
`https://xstats-apps.12306.work/api/v1/update/check`. Failures fall back serially to the other endpoint,
and the first success stops further requests so one check is not reported twice. Requests use the stable
`XStats/<app-version> (macOS <system-version>)` User-Agent format for regional routing and diagnostics.
The original 32-byte random value is generated with `SecRandomCopyBytes` and remains in the app's local
`UserDefaults`; it is not a credential, is excluded from WebDAV settings sync, and never invokes Keychain
authorization. No hardware serial number is used. The response has the same release manifest fields previously read from the static appcast (version,
date, notes taken from `CHANGELOG.md` by `scripts/appcast.py`, zip URL, sha256 and size). An
update is installed only after: sha256 matches, the zip holds exactly one `.app`, its bundle ID and
version match, `SecStaticCodeCheckValidity` passes with a requirement pinned to the running app's
team, and `spctl --assess` accepts it (notarized). The old bundle is renamed into a same-volume
temporary folder, the new one moved into place (restored on failure; an administrator prompt is
used when the folder is not writable), and a detached shell waits for the process to exit before
reopening the app. After an update the old helper may still be running; the app unregisters an outdated
helper, re-registers the bundled version, and verifies the protocol before privileged calls resume.

`server/api` is one Go program. Fiber exposes `POST /api/v1/update/check`, authenticated
`PUT /api/v1/releases/current`, and the aggregate `GET /stats` dashboard. GORM uses
`github.com/libtnb/sqlite` with WAL and one database connection so concurrent checks cannot compete for
SQLite's single writer. Each installation row stores only the SHA-256 installation ID, current version,
first/last check times and check count; request IPs and monitoring data are not persisted. The release
endpoint requires `XSTATS_RELEASE_TOKEN`. `scripts/publish_release.sh` uploads the dmg and zip to
object storage with `mc`, verifies each one by re-reading it from the CDN, creates the GitHub Release
that carries the dmg for manual downloads, and only then submits the generated appcast through
`scripts/publish_api.py` to both regional services. The manifest lands last, so an installed app never
sees a version whose package is not yet in place.

## WebDAV settings sync

Account login, OAuth callbacks and the old account backend have been removed.
Sync only needs the user's WebDAV server and does not run at startup, on wake or on settings changes.

WebDAVSync uses Basic authentication over HTTPS and reads/writes xstats-settings.json in an
existing directory via GET/PUT. The temporary URLSession does not share cookies or credentials,
uses normal certificate validation, refuses redirects, and streams downloads with a 1 MB limit.
GET requires HTTP 200; PUT accepts 200, 201 or 204. Writes are not retried automatically.
See [WebDAV PUT semantics](https://www.rfc-editor.org/rfc/rfc4918#section-9.7).

SyncController runs on the main actor and allows one request at a time. Upload requires confirmation
and overwrites the remote file without merging. Download validates a SettingsBackup envelope
(format, version, timestamp and SettingsDocument schema), then waits for confirmation before applying
the existing per-field validation. Cancellation and failures do not change local preferences.

SettingsDocument remains an explicit allowlist. WebDAV credentials, connection settings, local page,
helper state, monitoring data and history are excluded. Directory URL and username are saved separately
in local defaults. Passwords use the file-based login Keychain service work.12306.xstats.webdav,
with a separate account hash for each endpoint and username. Password storage must succeed before
new connection details are committed. JSON backups are not additionally encrypted at rest.

Configure every Mac separately. This version does not create remote directories, poll in the background,
merge concurrent edits, or use iCloud. A missing remote file requires an initial upload.

## Power, disk and history

- System, adapter and battery power come from SMC `PSTR`, `PDTR`, `PPBR`. GPU power comes from the
  IOReport *Energy Model* `GPU Energy` counter; CPU energy counters on recent chips barely update,
  so CPU power is not shown. Cluster frequencies are residency-weighted averages of *CPU Complex
  Performance States*, using the `voltage-states*-sram` tables of `pmgr` (Hz on older chips, MHz on
  newer ones). Each IOReport sample costs about 5 ms of CPU, so it runs at most every 2 s and only
  while a page shows it.
- Disk activity diffs `IOBlockStorageDriver` statistics; SSD health reads the NVMe SMART log
  through the system `NVMeSMARTLib` plug-in (no root).
- `HistoryRecorder` folds each sample into a per-minute record (averages, CPU and temperature
  peaks, worst memory pressure) and writes it to `history.sqlite`; records older than 8 days are
  pruned hourly. Queries bucket by 1, 5 or 30 minutes and charts break lines across gaps.

## Localization

Source strings are Simplified Chinese. `scripts/l10n_wrap.py` wraps every Chinese literal in
`tr(...)` (skipping logger calls, `case` patterns and multi-line strings) and lists the keys.
`tr` returns the source for Simplified Chinese. English uses the Swift tables; Traditional Chinese,
Japanese, Korean, German, Spanish, French and Arabic load bundled TSV catalogs from Localization/Resources.
Exact keys take priority, followed by templates with {} or numbered placeholders. Captured strings
are translated recursively; inserted values are not reinterpreted as placeholders. Missing entries
fall back to English, then the source. Translation tables are immutable; global language and caches
share a lock because background collectors also call tr.

The language picker uses a searchable native popover, native language names, a selected checkmark,
and keyboard navigation. Search accepts Chinese, English, native names and language codes.
Automatic detection walks global AppleLanguages in preference order, distinguishing zh-Hans from
zh-Hant/TW/HK/MO and falling back to English if no supported language is found. Existing system,
chinese and english preference values and WebDAV settings remain readable.

Changing language rebuilds SwiftUI roots and AppKit menu/window titles. Each SwiftUI window and popover
receives the selected locale and layout direction; Arabic uses right-to-left layout and isolates
interpolated paths/numbers with Unicode directional isolates. Dates use L10n.locale. For remote geographic
data available only in Chinese/English, other languages use English names; country names use the locale.
Some macOS-provided names follow the application's AppleLanguages on the next launch.

The nine supported languages are zh-Hans, zh-Hant, ja, ko, en, de, es, fr and ar. Snapshots accept these
codes through --snapshot <dir> --language <code>, include the language picker, and record missing
translations in untranslated.txt. New interface strings must be added to all catalogs; coverage and
placeholder checks accompany the localization tests.

`--snapshot <dir>` renders every main-window page, popover, settings section and the menu bar in light and
dark, through real `NSHostingView`s in off-screen windows — `ImageRenderer` washes out pages
that contain bitmaps.