fix(translation): stop shipping blank strings for terms the protector covers whole

[0xra0]

What this fixes

quality_report_20260810_093056, over a real ru→uk starfield_ru.STRINGS.xml
run, reports 344 errors — 343 of them EMPTY_TRANSLATION (severity error:
"Will break in game"). Those rows exported as <Dest></Dest>, which the game
draws as nothing at all.

They come out of one chain in which every step is locally correct:

1. TermProtector matches a term and replaces it with [[TK_…]]. When the match
   covers the whole string, the model is handed a bare token.
2. The model returns the token untouched — all it was given.
3. restore_text puts the term back, so the result is byte-identical to the
   source by construction.
4. _is_untranslated_echo sees a verbatim copy carrying ы/э/ё/ъ, reads it as
   "the model never translated this", and _translate_single returns None.
5. The row is never filled, and a re-run repeats every step identically.

Neither subsystem is wrong on its own; they had no way to tell each other the
verbatim-ness was intended. 296 of the 343 blank rows.

The trigger

star_system_name was [Cyrillic]+(\s+[Cyrillic]+)*\s+[IVXLCDM]+(?:[-–—]\w+)? —
an unbounded run of Cyrillic words followed by anything spelled out of
I V X L C D M. A lone C or D qualifies, so it claimed:

string	what it actually is
Портативный CD-плеер	a portable CD player
Шкура ходока четырёхногих карпов C	an item name
Активировать вспомогательное питание блока D	a quest objective
Отправиться на Арктур II	a quest objective — the verb was swallowed too

6,051 rows — 12 % of the file — reached the model as a bare [[TK_…]] token.

Changes

File	Change
gui/term_protector.py	Pattern requires one capitalised Cyrillic word (internal hyphens kept, for Аль-Баттани), a well-formed Roman numeral rather than a letter set, and an optional single-letter body suffix (a–f are the only ones the shipped file uses).
gui/term_protector.py	star_system_name moved from STRUCTURAL_PATTERNS into SOFT_CATEGORIES. A designation is a proper noun, so "Protect named entities" must govern it — with it off, Брэдбери I reaches the model to be transliterated as Бредбері I.
gui/ollama_worker.py	New _protection_covers_everything / _is_protected_verbatim. A fully covered string restores and returns before the network call, and both echo guards exempt an intended passthrough.
gui/ollama_worker.py	_clean_translation's mid-length shrink floor gated on closely_related, as its two immediate neighbours already were.
bethesda_strings/xml_handler.py	write_sst_xml omits untranslated rows and returns the count written.
gui/main_window.py	Export reports what actually landed in the file, plus how many rows were left out.
gui/quality_checker.py	printf_var matched and counted case-sensitively.

Why the export change is the part that broke the game

An SST file is a patch dictionary: an entry claims a finished translation
for its sID. <Dest></Dest> does not read as "not translated yet" — it reads as
"translate this to the empty string".

Two things in this repo already had the right policy and were not consulted:

• parse_sst_xml skips entries with no <Dest>.
• apply_changes_to_file (the binary .strings path) falls back to the
  original text.

So the same project saved as .strings was fine and saved as .xml was not.
Omitting the row restores symmetry with the reader and makes write→parse
lossless.

The second blanking path

_clean_translation's garbage detection discarded correct, legitimately shorter
East-Slavic output — «Сопротивление» → «Опір» (13 → 4) and «Стрельбище» → «Тир»
(10 → 3) both cleaned to "". The text_len == 1 check still catches a
genuinely empty reply.

The lone MISSING_TAG was a phantom

_COMPILED_PATTERNS compiled every entry with IGNORECASE, so the printf
conversion class also accepted U, and .MG%UYOL;{M&:AF — Starfield's
deliberately-scrambled terminal text — registered as a missing %u. In C the
case is the conversion: %U is not one, and %x/%X print differently. Every
other pattern keeps IGNORECASE, where <BR> and </FONT> are real spellings.
3 phantom matches over the shipped file (%U ×2, %O ×1).

Measured

	before	after
whole-file tokenisation (named entities off)	7,748 rows	1,697
of the 104 distinct broken strings, still fully covered	88	0
re-export of the same project: entries with a blank <Dest>	343	0

All 343 previously-blank rows now either reach the model with real text or return
verbatim without a call.

Tests

1083 passed, including 56 new across test_star_system_protection.py,
test_protected_passthrough.py, test_sst_export_untranslated.py and
test_printf_tag_case.py.

tests/test_term_protector_obfuscated_codes.py had already recorded a symptom of
this over-match as an accepted limitation — star_system_name claiming код C
so obf_code lost the trailing $1000. It failed by succeeding; the test now
pins the fixed behaviour.

Out of scope

UI_OVERFLOW (107 warnings) is a visual finding and was deliberately excluded.

One thing worth flagging separately: string 45533's translation is genuinely
truncated (826 → 230 chars, 15 sentences → 4, ending in a stray [). Its
MISSING_TAG was the phantom tag fixed here, but the content loss is real and
the checker reports it only at info severity, because SUSPICIOUSLY_SHORT
fires below 0.20× and this is 0.28×. Not addressed here — it affects every row in
the file and was not in the critical tier.

🤖 Generated with Claude Code