Add nine languages: Polish, Czech, Greek, Hebrew, Turkish, Bulgarian, Italian, Portuguese (PT + BR)

[vstrelnikof]

Closes #2.

Fifteen bundled layouts instead of six. Each language is a layout TOML
plus a full dictionary, so the detection that makes uk-UA ↔ en-US
reliable now applies to cs-CZ, tr-TR, it-IT and the rest. The
active-layout filter is untouched — a two-keyboard user still loads two.

Layout mappings were generated from xkeyboard-config rather than
transcribed from keyboard diagrams. The generator reproduced the
existing de_de.toml byte-for-byte as a control before any new file
was kept; the technique is written up in docs/ADDING_A_LANGUAGE.md
for the next person.

Three bugs found on the way

All three produced a green run, which is why they had survived.

Polish and Greek would have shipped as mojibake. Hunspell declares
a dictionary pair's encoding once, in the .aff — the .dic has no
SET line of its own. We scanned each file separately and fell back to
Latin-1 when there wasn't one, so every .dic hit that fallback. That
was invisible for the six languages bundled so far (four are UTF-8, and
German genuinely is Latin-1), and fatal for anything else: słowo
landed in the FST as s³owo, which matches no lookup and looks enough
like a word to survive a skim. Fixed by reading the .aff once and
decoding both halves with it, with real 96-entry tables for ISO-8859-2
and ISO-8859-7 generated from Python's codecs rather than typed by hand.
An unknown or absent SET is now an error — guessing is the failure
mode here, not a convenience. Polish went 2.28 M → 3.96 M forms once
the affix conditions actually matched.

The French source had been 404ing. Upstream moved it from fr_FR/
to fr_FR/dictionaries/; the fetch printed one line on stderr and
exited 0, leaving the stale wordlist in place. Fixed, and the command
now tallies failures and exits non-zero. The refresh gains 11 701 forms
and loses 26, all elision artifacts like d'pick-up.

FLAG num was rejected outright on the grounds that no bundled
dictionary used it. Turkish does.

Regenerating also turned up a UTF-8 BOM on pt_BR.aff, previously
invisible because valid-UTF-8 files short-circuited before the SET
scan.

Two caveats worth reading before merge

Polish maps to the same characters as US English. The layout
essentially every Polish user has enabled is the "programmer's" one —
QWERTY with every diacritic on AltGr, which PolterType doesn't track.
So there is no pl-PL ↔ en-US mistake to make and none to correct. The
file still earns its place: it gives Polish an id the OS switcher can
target, and it hangs the Polish wordlist off a layout, which is what
stops Polish prose being dragged toward whichever other layout is
active. Stated in the TOML header, CREDITS.md and the changelog
rather than left to be discovered.

Hebrew's Hspell is AGPL-3.0-or-later — the strictest licence in the
tree and the only one with a network clause. It attaches to the data,
not the engine, and we build no network service, so a locally-installed
desktop app is within terms; but a downstream packager needs to know.
Checking it prompted an audit of all thirteen against their upstream
licence files, which found the old blanket hedge ("most often
GPL-2-or-later or LGPL/MPL") was also wrong about Russian — it's BSD.
CREDITS.md now carries a per-language table.

Cost

+73 MB in the repo, +57 MB in every installer — 30 MB of that is
Turkish alone, whose agglutination expands to 5.8 M forms and two 15 MB
FSTs. That is the price of the detection quality, and nothing is read at
runtime without the matching keyboard enabled.

Hebrew would have been 141 MB expanded faithfully — its affix table is
3335 prefix rules and no suffix rules, encoding the clitic particles as
affixes — so it ships stems only via a new ExpandMode. Its script
already separates it from everything else bundled, so the dictionary is
a refinement there rather than the load-bearing signal.

Verification

• 390 tests pass; clippy --all-features -D warnings, fmt --check
  and cargo deny check all clean.
• New tests cover the codepage decoders, the BOM, FLAG num, the
  distinctive keys of every new layout, and — the one that would have
  caught the mojibake — that each bundled dictionary contains ordinary
  words of its own including diacritics, since the ASCII-only words
  were correct the whole time.
• A test now pins build.rs::LAYOUTS against BUNDLED_LAYOUT_STEMS;
  that drift previously only produced a startup warning.
• Ran the app on this machine: all fifteen are registered, the two
  active in the OS load, the other thirteen are skipped cleanly with no
  missing-TOML warnings.

docs/ADDING_A_LANGUAGE.md gains the two steps it never had — the WiX
entries (skip them and the language works everywhere except Windows,
silently) and registering the id with the OS switchers.