v3.4.0

[minirang]

Description

• Fixed: an incomplete statement was reported on the line after it. A statement that stopped early (set number x to followed by another line) pointed the caret at the next statement, and a block with no end pointed at an empty line past the end of the file. Errors now sit on the statement's own line, at the point where it ends, and a missing end says where the block was opened.

• Fixed: errors from imported files were drawn from the wrong file. An error inside a module's function or top-level code used the module's line and column but showed the importing file's line, and an error in a function declared by an earlier cuffsh entry showed the current entry. An error now carries the text it points into, so the report shows the right line and names the file (Runtime Error in './lib/ratio.cuff' at line 3). A module that fails to parse (E5-002) shows the use statement and then the module's own line and column.

• Fixed: an error inside an f-string's braces pointed at the start of the f-string. In print(f"fib(10) = {fib(10)}") a failing call put the caret under the f, because everything in the braces was placed at the literal's location. The error is now reported at the call itself, also after escapes such as \t or \", after Korean text, and when the literal runs over several lines.

• Behavior change: columns count characters, and the caret follows display width. column used to count bytes, so a Korean string earlier on a line made the headline disagree with the caret. It now counts characters. The caret copies tabs and gives wide characters two columns, so it lines up in a terminal. Offsets are still bytes.

• Fixed: a UTF-8 byte order mark at the start of a file was rejected by cuffc. The engine now ignores it, as the shell already did.

• Behavior change: text left over inside an f-string's braces is an error. f"{n 2}" used to evaluate n and silently ignore 2; it is now E2-001, and errors in an f-string expression name the expression.

• Fixed: a program the parser accepts could still fail, or crash a small thread. A chain of operators, indexes or slices just under the expression limit failed with E4-017 because evaluating it used stack in proportion to its length. Such chains are now evaluated iteratively, and a long expression tree is freed iteratively too (it overflowed a 256 KiB stack when it was merely destroyed). A stack budget configured larger than the thread's real stack used to turn runaway recursion into a crash; it is now reduced to the stack that exists. The relationships between limits are checked at compile time (static_asserts in Limits.h) and covered by new tests.

• Fixed: one Session per thread was a data race. The table of identifier names shared by all threads was unsynchronised; it is now locked. A thread-safety test runs under ThreadSanitizer in the documented setup.

• Fixed: the Windows build failed with MinGW-w64 from MSYS2. GetCurrentThreadStackLimits is a Windows 8 function that headers declare only for a newer _WIN32_WINNT than this project targets, so Attributes.h did not compile there (older MinGW headers let it through). It is now looked up when first needed, which works with any header and on a Windows without it, where the engine uses its default stack budget.

• Fixed: two shell tests failed on macOS. The test harness took a closed pseudo-terminal to mean "still running" because macOS reports it as end-of-file where Linux raises an error, and the fuzzer wrote to a shell that had already exited (macOS refuses such a write). Both are handled; the shell itself did not change.

• Fixed: calls queued by async survived an error. When a queued call failed, the calls still waiting ran at the next cuffsh entry. The queue is now emptied when an error ends a run.

• Behavior change: cuffc rejects unknown options and more than one script. An unrecognised option used to be taken as the script's file name. --help now writes to standard output, and a file is read as bytes on every platform (a text-mode read on Windows changed line endings and stopped at a Ctrl+Z byte).

• The shell follows the engine. cuffsh no longer parses, runs or formats errors itself: it hands each entry to cuff::Session and prints the finished report, so its output is the same as cuffc's by construction. CuffEngine::Options is now the interpreter's own configuration, and both programs read the engine's command-line flags through one function (EngineOptionsCli.h). API change: CuffEngine::execute no longer returns the syntax tree it ran.

• Clearer syntax messages. Messages say end of line and end of file instead of a quoted '\n' or ''.

• Faster reading of large programs. The token list is reserved up front (about 9% faster tokenizing of a 3 MB program), and cuffc hands its source to the engine instead of copying it.

• Documentation is organised and checked. docs/ is now guide/, reference/, internals/ and contributing/, with a page for every library, cuffc, cuffsh, the built-in functions, the error catalogue, the limits and the architecture. Limits, error codes, the function index and the command-line help are generated from the code (python3 tools/docs.py generate), and python3 tools/docs.py check verifies links, library coverage, the version and that every runnable example prints what the page says. SPEC.md, REGEX.md, IMPLEMENTATION_NOTES.md and EXTENDING.md moved to docs/reference/language.md, docs/reference/patterns.md, docs/internals/design-notes.md and docs/contributing/extending.md.

• Tests and workflows. New unit tests for error locations, limits, and threads. cli/tests/requirements.txt lists the shell tests' Python packages. The shell workflow sets up a pinned Python instead of --break-system-packages, builds cuffc so the cuffsh-versus-cuffc comparison actually runs, prints tool versions, and also runs when the engine changes. run_tests.ps1 expected the old four-digit error codes and its -TimeScale option never reached the tests; both work now. Unit tests link with -pthread.