`docs/METHOD.md` asks each deviation to name the reference fact and why benilla differs. About 130 `Deviation:` comments gave no reason, and the Warden comment lost its reason when the refusal became a read-past.
- `scripts/genmap.sh --check`, now a gate, fails a `Deviation:` comment without "because" and a `Deviates` row without a Why.
- `scripts/deviation-why.allow` lists the 43 comments whose change recorded no reason. It only shrinks: an entry that gains its reason, or is new against main's copy, fails.
- About 190 comments that gave their reason in other words now say "because", with their meaning kept.
- Nine reasons, the Warden one among them, come back from the commits and pull requests that made them.
Tested by the gate on the full tree, and by its failure on a comment with no reason and on a stale entry.
Fixes#2856Fixes#2857🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ten fixes, each with a test that fails without it.
- Status lines are cut at 255 bytes (`0x41a049`), banners too (`0x419f19`), and an empty line is dropped without raising the severity (`0x41a05f`).
- A message frame's `bytes=N` holds N−1 bytes. A negative N empties its lines, since `0x771692` tests only zero.
- The reward panel after a progress panel plays the completion chime (`0xbe081c`, `0x500d36`).
- A panel opens only for a held unit, game object or item (`0x500c16`-`0x500c35`).
- Readable pages are rewritten into their language for the reader (`0x49b560`).
- A page of 1024 bytes or more reads empty, as `0x4191b0` does at capacity 0x400.
- World entry wipes only the four unsaved caches (`0x555740`)., keeping saved asks.
- Two comments cite the right addresses.
Fixes#2580Fixes#2690Fixes#2692Fixes#2694Fixes#2696Fixes#2700Fixes#2701Fixes#2710Fixes#2711Fixes#2712🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Lua numerals and the sscanf ports used Rust's correctly rounded parse. 1.12.1's C runtime rounds its own way: `__RoundMan` rounds up only when the round bit and some bit two past it are set (`0x40fb43`), so ties truncate and `6e23` reads one ulp low. It flushes results below 2^-1022 to +0 (`0x741ed3`) and keeps 24 digits (`0x41157f`). benilla-re holds these at proof two.
One port of that conversion now lives in benilla-lua (`strtod.rs`, `ld12.rs`). It serves `str2d` and `%f`, which rounds once to float (`0x40feef`) and consumes a digitless exponent marker (`0x7451a5`). `%d` wraps in 32 bits (`0x7456c9`). The time zone is read once per process (`0x414568`), and two client faults are written as Deviations.
A table of 3487 cases generated from benilla-re's transcription, and the exe's fixed cases, pin it.
Fixes#2647Fixes#2648Fixes#2649Fixes#2650Fixes#2651Fixes#2652🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A script error reached the `seterrorhandler` function once per frame, from a queue, after every other listener had run. In 1.12.1 the handler is the error function of the pcall each script runner makes (`0x704e4f`-`0x704e79`; the error function `0x703b40`). So it runs at the raise, inside the failing script: `this`, `event` and `arg1` are still the failing handler's, `debugstack()` shows the failing call, and the event's next listener has not run yet.
Every benilla runner now passes that error function: handlers, OnLoad, file chunks, RunScript, key bindings and slash commands. It returns the handler's result, as `0x703b40` does. The queue is gone.
A handler body that fails to compile reaches no handler (`0x704c70`). benilla's host log keeps each original fault, and `/errors` replays the ones raised before benilla's layer loaded.
Tests cover the raise-time globals, the stack, the listener order, OnUpdate and chunks.
Fixes#2656🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Lua 5.0 behaviours of the client that an addon can observe, which benilla-lua's 5.1 base got differently:
- string.sub and string.find keep a position's low 32 bits (`0x7fbd55`).
- An escaped byte at or above 0x80 matches nothing (`0x7fcbc6`).
- `error()` checks its argument (`0x702993`) and adds the position at any level but 0 (`0x7029a8`).
- `__unm` gets the operand and nil (`0x6f8edc`).
- `__index` chains reach 102 tables and `__newindex` 101.
- math.deg multiplies by 180/pi (`0x7fb14d`).
The bare aliases run as their own `compat.lua` chunk (`0x703b21`), each on the client's line. Its closures keep a local `math`, so their error lines match and replacing the global `math` leaves them working.
`seterrorhandler` keeps the top value, and nil clears it (`0x702930`).
Each fix has a test.
Fixes#2632Fixes#2633Fixes#2634Fixes#2635Fixes#2636Fixes#2637Fixes#2638Fixes#2639Fixes#2640Fixes#2654🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On Windows, `vm_parity`'s cos sweep failed on one item, `cos(1e300)`, whose last digit was the host libm's. 1.12.1 never computes it. `math_sin`, `math_cos` and `math_tan` are a bare `fsin`, `fcos` and `fptan` (`0x7fae1d`, `0x7fae4d`, `0x7fae7d`), and a finite operand of 2^63 or more is outside the x87's range and is left unchanged:
- `math.sin(1e300)` and `math.cos(1e300)` answer `1e300`;
- `math.tan` answers `-1.#IND`: `fptan` pushes nothing, so the pop at `0x7fae84` takes the operand and `0x7fae86` stores from an empty stack.
benilla-lua now answers the same. The sweep's recorded answers for that argument are the reference's, the same on every platform.
Test: the parity sweep's `1e300` items for `math.sin`, `math.cos`, `math.tan` and their degree forms.
Fixes#1999🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ten differences from 1.12.1 in the UI core.
- StatusBar range and value follow `0x783270`/`0x7832d0` (a reversed pair collapses, a value before a range is dropped); the health watchers set both each update; the fill draws once both are set, its coordinates built from scratch (`0x7833c0`).
- A full collection runs before PLAYER_LOGIN (`0x490941`); the addon harness runs the client's world-entry cascade.
- Every index argument raises its own usage line (`0x6f34d0`, then the binding's literal).
- Strata, layer and frame-type names match whole (`0x6ee32c`); stack lines name unnamed functions by global (`0x6fbf90`).
- SetCursor takes a custom file (`0x523c40`); EditBox accepts OnInputLanguageChanged (`0x77a40d`).
- debugprofilestop answers milliseconds (a Deviation: multi-core 1.12.1 returns cycles).
A byte audit re-derived every address.
Fixes#1178Fixes#1915Fixes#1917Fixes#1908Fixes#1839Fixes#1869Fixes#1170Fixes#1105Fixes#1186Fixes#1952🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Six places where benilla's Lua answered unlike 1.12.1's:
- The lexer took `[=[` long-string levels; 1.12.1's `luaX_lex` tests one byte after `[` (`0x6ff793`) and `--[=[` is a line comment (`0x6ff6c9`).
- `math.random` ran a glibc-style `rand` seeded with 1; now MSVC's (`0x7400e5`, `% 0x7fff` times 1/32767 at `0x7fb310`), clock-seeded as the client is (`0x5d1c8b`), fixed in captures and the addon harness.
- `getfenv`/`setfenv` honour a raw `__fenv` field (`0x702ba0`), and `setfenv` returns nothing.
- An object's text is `type: %08X` in the embedding API too (`0x74301d`).
- `time` converts a date table through MSVC's `_mktime` (`0x74066b`).
- `date` formats with MSVC's C-locale patterns (`0x8323a0`): `%c` is `09/09/01 01:46:40`, `#` drops zeros.
Each has a test that fails without it; the addon corpus parses as before.
Fixes#1257Fixes#1258Fixes#1510Fixes#1511Fixes#1503Fixes#1504🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
benilla-lua's tables and globals now follow the client's Lua 5.0, not 5.1.
Tables:
- They size, fill and rehash as 5.0's ltable.c does: the hash part doubles (`0x6fab90`), and a log size of 0 is the shared dummy (`0x6fa601`).
- A hash part past 2^24 nodes raises `table overflow`.
- Constructors store every 32 items.
- A weak table's clearing removes the key, so `next` from it raises (`0x6f79d0`).
- A weak-keyed table keeps its string keys (`0x6f79f0`).
Globals and libraries:
- `ipairs` returns itself, and a second argument steps once.
- `setglobal` assigns the top of the stack.
- The state opens in the client's order, with no `_G`.
Each new test fails without its fix.
Fixes#1487Fixes#1789Fixes#1790Fixes#1169🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The collector now paces as the client's Lua 5.0 does.
The threshold:
- The state opens at 4× its count (`0x6f6ec8`), and InitLua then sets 256 KB (`0x7039f2`).
- After a collection it is the count plus the larger of the old headroom and a quarter of the count (`0x6fae00`). The headroom is capped at the free space in the client's six block pools, which a new book models (`0x6fae90`).
- The memory watch collects from a threshold of 0 (`0x703bd6`), and varargs never collect.
The stack and call records start at 45 and 8 slots and grow by 5.0's rule (`0x6f5f60`). They are checked once per collection, during mark, and halved under a quarter's use (`0x6f7949`). A tail call checks the stack where the call is made (`0x6f922c`).
A collection of a 10 MB heap takes about 1 ms; each change fails its test without it.
Fixes#1783Fixes#1784🤖 Generated with [Claude Code](https://claude.com/claude-code)
Lua memory now counts as the client's does:
- Every object is charged at its 32-bit Lua 5.0 size: a string is 17 bytes plus its length (`0x6f9da1`), a table header 32 (`0x6fa500`), a node 40, and a closure 36 plus 4 per upvalue (`0x6f9ebf`).
- Prototypes count 5.0's instructions and constants (`0x6fcb2b`).
- The string table and the concat buffer are counted as 5.0 sizes them.
The stock UI reads 10,618 KB, against 13,294 KB before.
The host's old C-host shim overrode benilla-lua's natives. It is gone:
- `collectgarbage` now moves the real collector, taking `__ftol`'s low word (`0x6f4400`);
- `gcinfo` reports that collector;
- `^` is the client's `math_pow` (`0x7fb4bd`);
- `loadstring` returns one value (`0x7032ba`).
An audit re-derived every size from the bytes. Each fix fails its test without it.
Fixes#1779Fixes#1780🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
benilla-lua's lexer and compiler now behave as the client's 5.0:
- A control byte is "invalid control char" (`0x6ffabb`).
- Every chunk lexes as source, binary ones too (`0x6f6a15`).
- Lines end at `\n` alone (`0x6ffb94`); a dot is never a NUL (`0x6ff98a`).
- A `while` condition over 100 instructions is too complex (`0x6fd03b`), counted as 5.0's code generator counts it.
- Upvalues get no name in errors (`0x6fc5e0`).
- `string.dump` refuses closures with upvalues, and dumps the stack top (`0x6f438a`, `0x6f4373`).
The client installs no Lua hook (`lua_sethook` has no caller), so the load's runaway guard has no client mechanism to follow. It is now an abort no `pcall` can catch, recorded as a deviation.
An audit re-derived every commit from the bytes. Each fix fails its test without it.
Fixes#1755Fixes#1756Fixes#1757Fixes#1758Fixes#1759Fixes#1767Fixes#1434Fixes#1445Fixes#1777Fixes#1778🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Dev builds get three tools, all behind the `dev` feature:
- **Hot reload:** an edited addon `.lua` is swapped into the closures already running, keeping state and event registrations. Edited XML handlers and benilla's own `assets/ui` layer reload too.
- **An addon profiler:** Lua time and heap per addon.
- **A Lua inspector:** tables, frame scripts and upvalues, with scalars editable in place.
benilla-mlua's `unsafe raw_vm` becomes a safe lent `with_vm`.
Ours on the branch:
- the top level is compared off the compile path, in a scratch heap;
- watched files are polled in slices, and each addon's XML is walked once per session;
- no lend while a callback of the state runs;
- profiler owners are cached per source epoch;
- key-ups reach the UI while an egui field has focus.
The player build carries none of it. Tests: the full gate chain on the merged tree.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Sam <samuv.eth@gmail.com>
Three tests failed on a Windows checkout:
- one split a source path on `/`;
- one split its own `include_str!` on `"\n}\n"`, which a CRLF checkout lacks;
- one captured a traceback's addon folder with a greedy `.*`.
Each now normalises the separator or line end, or bounds the capture. A sweep found the same path forms in four more source-walking tests (`fire_scan`, `world_api_wall`, `reference_strings` and `front_end`), now normalised too.
`.gitattributes` pins Rust sources to LF and leaves `third_party/kira`'s CRLF files as they are.
Built and tested on macOS. The Windows cases are reasoned from the path and line-ending forms, not run.
Fixes#798🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The table library and `unpack` now run as the client's 5.0 natives (`0x822d40`). benilla's closure layer over them is gone. `unpack` refuses past 2,048 slots (`0x6f2f43`) and reads its table alone. `getn` converts a numeric-string `n` (`0x6f4f80`). `next` says "for" (`0x871aac`).
The bit library answers unsigned except `arshift` and `mod` (`0x7faa6d`). Its folds take every argument, and `math.pow`, `atan2` and `mod` check argument 1 first.
A Rust callback's error now carries the caller's `file:line`, as `luaL_error` does (`0x6f4940`); only the sound verbs raise bare. Error text keeps its bytes, from `assert` to the error handler.
An audit re-derived every commit from the bytes. Each fix fails its test without it.
Fixes#1435Fixes#1448Fixes#1449Fixes#1480Fixes#1482Fixes#1483Fixes#1484Fixes#1485Fixes#1760Fixes#1761Fixes#1762Fixes#1763Fixes#1764Fixes#1765Fixes#1766Fixes#1771🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
benilla-lua's string and number library now answers as the client's:
- `string.byte` returns one byte (`0x7fc210`).
- `gsub` takes only a string or a function (`0x7fd163`), and `%0` raises (`0x7fce20`).
- `lower` and `upper` case-map UTF-8 (`0x7fbeb0`).
- `tonumber` with a base is MSVC's 32-bit `strtoul` (`0x40db09`), on every platform.
- Chunk names cut at a NUL and at `\n` (`0x835144`).
- `string.format` is the client's native one (`0x7fd3d0`): `__ftol` conversions, `%q`'s table, one-digit `n$`, and the client's errors.
- Number text is the MSVC CRT's, ported down to `$I10_OUTPUT`: `1.#INF`, `-1.#IND`, `1e+020`.
- A NaN made from numbers is the x87 indefinite, and the math library returns the runtime's own answers at infinities.
A fuzz against the client's own `sprintf`, run in emulation, matches 1.5M cases. Every fix fails its test without it. `tostring` and `%g` got faster.
Fixes#1367Fixes#1368Fixes#1373Fixes#1375Fixes#1376Fixes#1450Fixes#1481Fixes#1486Fixes#1770🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
benilla-lua now matches the client's Lua 5.0 wherever an addon can see the difference:
- Number keys hash as `hashnum` does (`0x6fa260`), so `pairs` walks them in the client's order.
- Calls nest at most 4096 deep (`0x6f61c0`).
- `^` calls the global `__pow` (`0x6f9b5e`).
- The parser's limits and messages are the client's (`0x6ff440`); the client has no assignment limit.
- A numeric for's variable is its counter (`0x6fe380`, `0x6f9262`).
- A generic for uses `forlist`'s layout (`0x6fe540`): the first name is the control, and closures share the variables.
- Only calls name a callee (`0x6fc499`), and a stray token reports `<eof> expected` (`0x871cf4`).
Each fix has a test that fails without it. Two audits re-derived every address from the bytes. The bench is faster or level.
Fixes#1369Fixes#1370Fixes#1371Fixes#1372Fixes#1374Fixes#1377Fixes#1752Fixes#1753Fixes#1754🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The script host now runs on benilla-lua alone. The vendored C Lua 5.1.5 (`third_party/lua-src`) and mlua are removed, about 17k lines. `vm_parity` checks benilla-lua against the answers the C build recorded on Linux and macOS, and `front_end` checks every corpus chunk's compiled shape against the C build's digest. `GRAMMAR.md` records each grammar change from 5.1, by client address.
A handler's `this`, `event` and `arg1`… are bound in one VM call. An event fire saves and restores `event`, while the handler runner saves only `this` (`0x704d50`).
Ours on the branch: VM fixes from review (hook dispatch counts, registers and frames around a collection, protected-call handlers, library errors) and `vm_parity`'s two-VM leftovers removed.
Tests: the full gate chain on the merged tree, `vm_parity` and `front_end` among them. Live: smoke and a stock-glue login. Load time is level with C Lua, and memory is within 7%.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Sam <samuv.eth@gmail.com>
benilla-lua is a Lua 5.1.5 interpreter in Rust that speaks the 1.12.1 client's grammar. It is opt-in behind the `benilla-vm` feature, with `benilla-mlua`, the slice of mlua that benilla-ui uses, over it. The C build stays the default. Its front end ports 5.1.5's `lparser.c` and `lcode.c`, so registers, limits and messages match the C build's.
Our finish on top:
- the collector counts growth as C does, off the write path;
- frames under a leaky generic `for` start nil, as `luaD_precall`;
- the budget hook counts what C dispatches, and costs nothing when unset;
- `__index` chains stop at 100;
- mlua 0.11's conversions;
- `vm_parity` answers recorded per platform.
Tests: `vm_parity` on both VMs; the front-end corpus (8,543 chunks) against `string.dump`; 7,613 app and UI tests on benilla-lua. Built on macOS; Linux by CI.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Sam <samuv.eth@gmail.com>
Sam Sung