v1.5.0

[minirang]

Description

• Operators are now enums instead of strings. BinaryOp/UnaryOp stored the operator as a std::string and the interpreter dispatched through an if (op == "is") ... else if (op == "+") chain. Profiling showed this cost 16.2 million string comparisons on a 635k-call benchmark — about 25 per call, the largest single cost in the engine. Now a switch over BinOp/UnOp assigned at parse time. fib(27): 0.241s → 0.149s.
• Variable and function names are interned to integer IDs (engine/common/NameInterner.h), so scope lookups compare uint32_ts over a contiguous vector instead of strings. A 400-global lookup benchmark went 0.059s → 0.027s. Names are still stored alongside for error messages.
• Interpreter scope-storage optimization. Profiling showed parameter binding dominated call cost (~60ns per parameter, 72% of a 3-argument call) because each one was an unordered_map insert: a hash plus a node allocation. Scopes are small in practice, so Environment now stores variables in a contiguous vector with linear lookup — one allocation per scope instead of one per variable — falling back to a lazily-built index once a scope exceeds 16 entries, which keeps large global scopes fast. Measured: 0-argument call overhead 72ns → 32ns, 3-argument 252ns → 164ns.
• Decided against the larger slot-resolution refactor (pre-resolving every variable to an array index at parse time) and a bytecode VM for now. The measurements above captured most of the available win for a fraction of the risk, and the engine is now in the same performance range as CPython on call-heavy code. See docs/IMPLEMENTATION_NOTES.md section 21 for details.
• Net effect of optimizations this release: fib(27) 0.32s → 0.140s (~56% faster), a 2M-iteration loop 0.287s → 0.173s, 3-argument call overhead 252ns → ~145ns.
• Added DLC:json (to_json, from_json). Object/array/string/number/true/false/null map onto map/list/str/number/boolean/empty. to_json(value, indent) pretty-prints. The parser is strict per RFC 8259 — trailing commas, single quotes, unquoted keys, leading zeros, NaN/Infinity, and trailing content are rejected — and decodes \uXXXX escapes including surrogate pairs. See docs/IMPLEMENTATION_NOTES.md section 19.
• Error-code audit and normalization. ArgumentError (E4010 ArgumentCountMismatch) was being used both for wrong argument counts and for bad argument values (sqrt(-1), to_number("abc"), malformed JSON). Added ValueError / InvalidArgumentValue (E4025) and rerouted the six value-problem sites to it. Also normalized message capitalization across every throw site: all messages now start lowercase, since they're always printed after a ...at line N, column M: prefix.
• The regex matcher is now codepoint-based instead of byte-based. [any] matches one character rather than one byte ("안녕하세요" is "[any]5" is now true, previously false); literal non-ASCII characters in a pattern compile to a single multi-byte literal node; negated sets ([!num]) match non-ASCII codepoints; search() only starts attempts on codepoint boundaries and [one:...] alternatives must end on one. [edge] treats non-ASCII letters as word characters, per REGEX.md section 17. [let]/[str]/[word]/[num]/[hex] stay ASCII-only per spec. See docs/IMPLEMENTATION_NOTES.md section 17.
• Fractional indices and range bounds are now a runtime error (FractionalIndex, E4024) instead of being silently rounded. Applies to list[i], str[i], slices, and loop repeat bounds. Whole-valued doubles (2.0, 6 / 2) still work.
• Moved Utf8.h from engine/interpreter/ to engine/common/, since the regex engine now uses it too.
• Added tests/unit/ for standalone C++ unit tests, and moved the regex engine's own test suite there (78 cases, including new Unicode ones). tests/run.sh now builds and runs it first.
• Expanded integration tests: added tests/cases/json.cuff (plus 7 error cases), tests/cases/regex_unicode.cuff, tests/cases/whole_number_indices.cuff, and 3 fractional-index error cases. All 66 script tests and 78 regex unit tests pass unchanged.

---

This discussion was created from the release v1.5.0.