Fix CI for Pull Requests affecting Rust code

[moritz-gross]

@NSoiffer all PRs that touch Rust code seem to fail, see for example https://github.com/daisy/MathCAT/actions/runs/35412550935/job/105815034881?pr=776

test Languages::zh::tw::units::without_prefix_powers_of_2 ... ok
test Languages::zh::tw::units::without_prefix_other ... ok
test Languages::zh::tw::units::without_prefix_time ... ok

failures:

---- Languages::en::alphabets::cap_cyrillic stdout ----
Error: MathCAT crash! Please report the following information: 'MathCAT crash! Please report the following information: 'MathCAT crash! Please report the following information: 'RefCell already borrowed' at src/speech.rs:2811:0
' at tests/common/mod.rs:56:0' at tests/common/mod.rs:61:0

failures:
Languages::en::alphabets::cap_cyrillic

test result: FAILED. 5105 passed; 1 failed; 33 ignored; 0 measured; 0 filtered out; finished in 194.14s

error: test failed, to rerun pass --test languages
Error: Process completed with exit code 101.

cargo test also passes locally. My guess is that it relates to the (virtual) environment / filesystem for the rules files ?!?

[yasumorishima]

Three observations from hitting this on #779, in case they narrow it.

It is not limited to PRs that change Rust logic. #779 changes six Rules/Languages/ja/*.yaml files and tests/Languages/ja/ja.rs (new test cases only), and fails the same way: https://github.com/daisy/MathCAT/actions/runs/35448526023/job/105911688298

The failing test is not always the same one. It is Languages::en::alphabets::cap_cyrillic on #779 and on the run you linked, but on run 35410037310 (branch ja) the same panic came out of Languages::zh::tw::alphabets::cap_cyrillic. And in a single run, Test (no-unsafe, Rules.zip) fails while Test (no-unsafe, Rules-minimized.zip) passes. So it looks like shared state plus ordering, rather than anything in one language's rules.

The line you pasted is suggestive. On main (49de6d67), src/speech.rs:2811 is rules.unicode_full.borrow_mut().clear();, inside replace_single_char. Reading that function: line 2801 binds unicode = rules.unicode_short.borrow(), line 2824 rebinds it to rules.unicode_full.borrow(), and that Ref is still alive at lines 2844-2850, where the replacements are expanded and self.replace(...) is called. replace can re-enter replace_chars → replace_single_char (e.g. via speech.rs:578). If the inner call then needs a character that is not in the short table and is_file_up_to_date(...) at line 2809 answers false, it reaches line 2811 and takes borrow_mut() on the cell the outer frame is still holding shared.

If that is the path, the trigger is the rule files' timestamps rather than their contents, which would fit both "passes locally" and the Rules.zip / Rules-minimized.zip split within one run — and cap_cyrillic is a natural place to hit it, since capital Cyrillic goes through the full table and expands into a further replacement.

That last part is a reading of the code, not something I have run, so please treat it as a starting point rather than a diagnosis.