refactor(i18n): one file per locale with keys typed against English; add Italian - #120

The last open piece of the markdown viewer: read → edit → save now
happens where the reading happens, instead of needing an editor pane to
change three characters.

The interesting design is not the textarea, it is the write path. This
is the first renderer→disk write in this surface, the renderer holds the
backing path in its store, and markdown content is explicitly untrusted
— so a renderer-supplied write path is treated as attacker-controlled.
markdown-grants.ts keeps a per-window set of paths that a *native
dialog* or an *authenticated pipe client* opened this session, and
refuses to write anywhere else. MARKDOWN_READ_FILE is deliberately NOT a
grant source: it takes a renderer-supplied path, so granting on it would
let the renderer mint its own grants and the set would mean nothing. The
cost is that a dragged-in file saves through Save As the first time —
which is the user confirming the destination, i.e. the point.

writeMarkdownFile re-applies every read guard on the way out (extension
whitelist, 5 MB cap, symlink refusal — a grant must not be launderable
into writing a .ps1), adds optimistic concurrency on mtime, and renames
a temp file into place, because a half-written spec is worse than an
unsaved one.

Conflicts are detected by re-stat on focus rather than a watcher per
pane: it needs no lifecycle and covers the real case, an agent rewriting
the file while the user was in another pane. With a dirty buffer the
save is blocked and the banner offers overwrite / reload / save a copy.

The editor is a plain textarea — CodeMirror would add 1–3 MB for a
surface whose job is "fix a typo, add a bullet" — with Tab indenting
instead of moving focus, and Enter continuing list/quote prefixes
(ending the list on an empty item, as every editor does).

Unsaved buffers get a `•` on the tab and confirm before closing, through
the same request/pending/confirm split ConfirmCloseDialog already uses
for workspaces, so `wmux close-surface` never blocks on a modal.

Also closes the two remaining "worth fixing" items from the issue:
`--title` on `markdown set` (every CLI-pushed surface was labelled
"Markdown") and `wmux markdown get`, which lets an agent verify what it
actually pushed.

PR #120's typed keys earned themselves immediately: every one of the 15
new i18n keys was a compile error until it was declared in en.ts. One
hole found and closed while adding them — the toolbar's menu builder had
typed `t` as `(key: any)`, which exempted its labels from that check.

Closes #116

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MNZmKeLgMpkhvZtLFrm6kG

Second language added under the per-locale structure from #120, and the
two-edit claim holds: a locales/es.ts file plus one row in REGISTRY.
Language, SUPPORTED_LANGUAGES, the Settings dropdown and the isLanguage()
persistence guard all derive from that table and picked up `es` on their
own; the only other change is the test that deliberately pins the shipped
language set, which fails by design when the set grows.

Coverage is complete: es 107/107 keys, so nothing falls back to English.

detectDefaultLanguage() now resolves es-*, so a fresh profile on a
Spanish-display Windows starts in Spanish instead of English (the issue
#56 case). Existing installs keep whatever they persisted.


Claude-Session: https://claude.ai/code/session_0137J8T5Ckm8GodkfkJc1MQi

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>