v3.2.0

[minirang]

Description

• New: several libraries in one use statement. use DLC:math, DLC:string, DLC:convert loads all three; one use per line still works and gives the same result. Every item in the list needs its own DLC: prefix (use DLC:math, string is a syntax error). A dangling comma and a forgotten comma (use DLC:math DLC:string) each get a specific message instead of a confusing unexpected token ':'. An unknown library in the middle of a list (E5-004) is pointed at directly. SPEC.md still requires a module-load command to fit on one line, so a list cannot be continued onto the next line. UseStmt now holds a list of {name, loc} for the DLC form.

• New: typed function parameters. set func add_num(number n1, number n2) do:. Each parameter is name or type name, with number, str, boolean, list, map or match. Typed and untyped parameters can be mixed. Passing a value of another type is a runtime error, ParameterTypeMismatch (E4-030), reported at the call site. For async functions the check runs before the call is queued, so the error points at the call. empty is accepted by every type, the same rule as set. A type keyword counts as a type only when a name follows it, so (number) and (number, x) remain valid untyped parameters. Functions with no typed parameter skip the check entirely.

• New: an interactive CuffScript shell, cuffsh. Implemented cuffsh to allow users to execute scripts in a REPL environment. Added command-line options for script execution, timeout settings, and network configurations. Created a fuzz testing script to identify potential crashes and runtime errors in the shell. Developed a harness for testing terminal interactions, including escape sequences and cursor movements. Implemented comprehensive test cases for multi-line editing, error handling, and input/output behavior, and added a shell script to run the test suite with dependencies.

• Changed: error codes print as E<category>-<number>. E4005 is now E4-005, E2001 is E2-001, and so on. The enum values in C++ are unchanged; only the printed tag is split, by errorCodeTag(). The dash keeps a tag unambiguous if a category ever needs two digits (E10-001). Capacity within a category is still 999 numbers. 158 test/example files (.expected_code files and note: E#### headers) and all current docs were converted. Entries in this changelog from earlier releases keep the old format, as a record of what was printed then.

• Fixed: errors inside an f-string {...} pointed at the wrong source line. The expression is tokenized separately, so its positions started over at line 1, column 1, and the source snippet added in the previous release then displayed line 1 of the file for any error inside an f-string. Locations for tokens in the fragment, and for any error thrown while tokenizing or parsing it, are now pinned to the f-string literal, so the snippet shows the line the f-string is on. The caret sits at the start of the literal, not at the exact column inside the braces. Two cases added to tests/unit/error_snippet_test.cpp.

• Fixed: tests/run.ps1 did not work on Windows PowerShell 5.1. It used ProcessStartInfo.ArgumentList, which does not exist on .NET Framework, so every test failed with a null-method error. Rewritten, and these were fixed on the way: output is decoded as UTF-8 (the OEM code page garbled Korean output and failed every diff); \r\n is normalized before comparing; stdin is closed so input() sees EOF instead of waiting for the timeout; g++ runs through the same helper rather than & g++ 2>, since under $ErrorActionPreference='Stop' PowerShell 5.1 turns any compiler warning on stderr into a terminating error, and the build timeout is 600 s instead of 10 s; unit tests link with -lws2_32 and an 8 MiB stack on Windows, as Makefile.win does; the binary is cuffc.exe on Windows and ./cuffc elsewhere, so it also runs under PowerShell 7 on Linux/macOS. Not executed on a real Windows machine (none available here).

• Release binary is now stripped. -s is added to Makefile and Makefile.win: 1.18 MB → 670 KB with no change in speed. It is skipped on macOS, where Apple's linker ignores the flag with a warning.

• Performance: measured, nothing else worth changing. Against the original engine: fib(30), a 3M-iteration loop, 1M user-function calls, string/list/map work, f-strings and a 66k-line script are all at parity (differences inside run-to-run noise). The only profiler available was gprof, and its profile was flat (evalExpr, evalBinaryOp, scope lookup). Tried and not adopted: profile-guided optimization (0–4%, inside noise, needs a two-step build), -O2 instead of -O3 (same), static libstdc++ (startup 2.2 → 1.7 ms but a 2.4× larger binary), and dropping -fstack-protector/_FORTIFY_SOURCE (about 0–8%, not worth weakening a sandboxing interpreter). A micro-suite (200k-element sort, list_unique, split into 150k parts, replace/find/count over 850 KB, a 50k-object JSON round trip, a 50k-key map) found no quadratic behavior.

• Comments in engine/ cut from 1,534 lines to 92. Of 657 comment blocks, about 90 short ones remain: lifetime and ordering invariants, platform quirks (windef.h macros, Emscripten's shadow stack, SIGPIPE, pthread_get_stackaddr_np), stack-safety limits, the reason behind non-obvious checks, and the public CuffEngine::Options fields. Longer explanations are in docs/IMPLEMENTATION_NOTES.md. Checked mechanically: a string/char-literal-aware scanner found the comments; the preprocessed output of main.cpp is identical before and after once whitespace is ignored; every string and char literal in engine/ is byte-identical; build and suite pass unchanged. main.cpp, wasm/bindings.cpp and tests/ were not touched.

• Docs. SPEC.md: typed parameters (§9), comma-separated use (§13), new §14 on the error-code format and categories. README.md / README_ko.md: the pure paragraph still said the restriction covers only the function's own body, which was wrong since the call-graph hardening in the previous release, and is now corrected; typed parameters and the use list are documented; the error example shows the caret and the new tag format; the directory tree now lists dlc/ and net/, which were missing. docs/IMPLEMENTATION_NOTES.md: new §35–39, plus the tag format in §20 and the f-string fix in §29. docs/EXTENDING.md: new §12–14 (parameter syntax, use syntax, comment policy). SECURITY.md, tests/README.md and examples/README.md updated; examples/08_dlc_libraries.cuff shows both use forms.

• Tests: 8 new checks (typed_parameters, use_multiple_dlcs, and the error cases parameter_type_mismatch, parameter_type_mismatch_async, use_multiple_unknown_dlc, use_multiple_missing_prefix, use_multiple_missing_comma, use_multiple_dangling_comma), and the f-string cases above. The suite is 124 checks, up from 116.

• Verification. Clean build, 124/124 passing. Also run clean: AddressSanitizer + UBSan over the whole suite, with a leak check on the new-feature scripts; libstdc++ debug mode (_GLIBCXX_DEBUG); -pedantic -Wall -Wextra -Werror; each header under engine/dlc, engine/interpreter, engine/parser and engine/common compiled standalone. The one known LeakSanitizer finding (the intentionally circular lists in value_semantics.cuff) is the same pre-existing one. Not done: nothing here was compiled or run on Windows or macOS, since no such toolchain exists in this environment. The new engine code is standard C++17 with no platform-specific API. run.ps1 is reviewed by hand only.

---

This discussion was created from the release v3.2.0.