ww/selfhost/test/smoke.combined.ww

// time — clocks, instants, durations. Mirrors Hare's lib/time
// (ref/hare/time/duration.ha, instant.ha, arithm.ha,
// +linux/functions.ha). Calendar / date / strftime / timezone /
// sleep live in separate Hare modules and graduate when callers /
// supporting stdlib arrive.
//
// `duration` is a NAMED alias of i64 (lib/math/random precedent
// at lib/math/random/random.ww:8); ww treats NAMED as a newtype,
// so cross-i64 arithmetic inside this module needs explicit casts.
// Hare's structural alias semantics let those casts vanish, but
// our type checker is strict.

package time;

@symbol("rt_syscall") fn syscall2(num: i64, a: i64, b: i64) i64;
@symbol("rt_abort")   fn abort(msg: str) void;

def SYS_CLOCK_GETTIME: i64 = 228;

// ref/hare/time/duration.ha:6. 290y representable range.
export type duration = i64;

// ref/hare/time/duration.ha:9-18. Plan-9 naming (lowercase)
// diverges from Hare's uppercase per project rule 4.
export def nanosecond:  duration = 1i64;
export def microsecond: duration = 1000i64;
export def millisecond: duration = 1000000i64;
export def second:      duration = 1000000000i64;

// ref/hare/time/instant.ha:9. (sec, nsec) pair — NOT POSIX struct
// timespec (which uses u32 nsec). Layout matches Linux's struct
// timespec on 64-bit (i64+i64) so we can pass &instant directly
// to clock_gettime.
export type instant = struct {
	sec:  i64,
	nsec: i64,
};

// ref/hare/time/+linux/functions.ha:84. First cut exposes only
// realtime and monotonic; Hare's process_cpu / thread_cpu / boot /
// realtime_alarm / boot_alarm / tai graduate when a caller needs
// them (CLAUDE.md rule 9 — Hare-fidelity, no premature surface).
export type clock = enum i32 {
	realtime  = 0,
	monotonic = 1,
};

// ref/hare/time/+linux/functions.ha:138. Hare's now() also aborts
// on impossible errnos. (instant | oserror) is deliberately not
// the return shape — EINVAL / EFAULT are programmer errors (bad
// clock id, bad ptr), and a 1-word-payload sum return walks into
// task #9's cgen-divergence trap.
export fn now(c: clock) instant = {
	let i: instant;
	let rc = syscall2(SYS_CLOCK_GETTIME, (c as i32): i64, (&i): i64);
	if (rc != 0i64) { abort("time.now: clock_gettime failed"); };
	return i;
};

// ref/hare/time/arithm.ha:9. Adds duration to instant. The
// negative-duration branch normalises nsec into [0, second).
export fn add(i: instant, x: duration) instant = {
	let r: instant;
	let xi: i64 = x: i64;
	let sec: i64 = second: i64;
	let nsec: i64 = nanosecond: i64;
	if (xi == 0i64) {
		r.sec  = i.sec;
		r.nsec = i.nsec;
		return r;
	};
	if (xi > 0i64) {
		r.sec  = i.sec + (i.nsec + xi) / sec;
		r.nsec = (i.nsec + xi) % sec;
		return r;
	};
	r.sec  = i.sec + (i.nsec + xi - sec + nsec) / sec;
	r.nsec = (i.nsec + (xi % sec) + sec) % sec;
	return r;
};

// ref/hare/time/arithm.ha:26. Returns duration from a to b.
// Sign convention: b - a.
export fn diff(a: instant, b: instant) duration = {
	let sec: i64 = second: i64;
	let v: i64 = ((b.sec - a.sec) * sec) + (b.nsec - a.nsec);
	return v: duration;
};

// ref/hare/time/arithm.ha:32. -1 if a < b, 0 if equal, +1 if a > b.
export fn compare(a: instant, b: instant) i8 = {
	if (a.sec  < b.sec)  { return -1i8; };
	if (a.sec  > b.sec)  { return  1i8; };
	if (a.nsec < b.nsec) { return -1i8; };
	if (a.nsec > b.nsec) { return  1i8; };
	return 0i8;
};

// os — process and filesystem facade. The body of each call lands
// either in libwwrt.a (rt_syscall trampoline) or libc bindings,
// depending on how the program was linked.

package os;

import time;

@symbol("rt_syscall") fn syscall0(num: nr) i64;
@symbol("rt_syscall") fn syscall1(num: nr, a: i64) i64;
@symbol("rt_syscall") fn syscall2(num: nr, a: i64, b: i64) i64;
@symbol("rt_syscall") fn syscall3(num: nr, a: i64, b: i64, c: i64) i64;
@symbol("rt_syscall") fn syscall4(num: nr, a: i64, b: i64, c: i64, d: i64) i64;

// alloc / free — runtime mmap-backed page allocator. Untyped:
// `alloc(n)` returns a `*void` and `free(p, n)` requires the byte
// count back because rt_free is munmap-based and doesn't track
// mapping sizes (the kernel needs the length to release the
// reservation).
//
// Diverges from Hare. Hare exposes `alloc` / `free` as typed
// language builtins (`alloc(value, cap)?` / `free(ptr)`) that the
// compiler lowers to rt::malloc/rt::free; ww has no such builtins,
// so the rt-symbol surface is exposed directly. Stdlib callers
// that need a typed allocation pattern wrap this with a cast plus
// a stored capacity (see [[strings.dup]], [[memio.dynamic]]).
//
// OOM: rt_alloc is a bare mmap(MAP_ANON|MAP_PRIVATE) wrapper with
// no error path. The raw Linux mmap syscall returns a negative
// errno cast to `*void` on failure (e.g. `(void*)-12` for ENOMEM);
// the `MAP_FAILED` (`(void*)-1`) value is a libc-wrapper convention
// that rt_alloc doesn't apply. Neither `== nil` nor `== (void*)-1`
// catches it; any deref of such a return faults. Today the stdlib
// does not check; OOM faults on first dereference. A typed
// fallible variant is a future task.
@symbol("rt_alloc") export fn alloc(n: u64) *void;
@symbol("rt_free")  export fn free(p: *void, n: u64) void;
@symbol("rt_abort") fn abort(msg: str) void;

// Hare-style runtime check. Caller passes a message that's printed
// to stderr before exit(1).
export fn assert(cond: bool, msg: str) void = {
	if (!cond) { abort(msg); };
};

// Linux amd64 syscall numbers. Internal to this module — passed as
// the first arg of syscall0..4 via libwwrt's rt_syscall trampoline.
// `nr` is the type so the call sites can't accidentally pass an
// arbitrary i64 (`syscall1(0i64, ...)` no longer typechecks).
type nr = enum i64 {
	READ        = 0,
	WRITE       = 1,
	OPEN        = 2,
	CLOSE       = 3,
	LSEEK       = 8,
	ACCESS      = 21,
	DUP2        = 33,
	GETPID      = 39,
	FORK        = 57,
	EXECVE      = 59,
	EXIT        = 60,
	WAIT4       = 61,
	MKDIR       = 83,
	RMDIR       = 84,
	UNLINK      = 87,
	GETCWD      = 79,
	GETDENTS64  = 217,
	NEWFSTATAT  = 262,
};

// open(2) flags. Linux values, matching <fcntl.h>. Hare names them
// `fs::flag::RDONLY` etc; we use the same leaf names so callers say
// `os.flag.RDONLY` and `os.flag.WRONLY | os.flag.CREATE`.
export type flag = enum i32 {
	RDONLY  = 0,
	WRONLY  = 1,
	RDWR    = 2,
	CREATE  = 64,	// 0x40
	EXCL    = 128,	// 0x80 — pair with CREATE to fail on existing path
	TRUNC   = 512,	// 0x200
};

// lseek(2) whence. Hare names it `io::whence`.
export type whence = enum i32 {
	SET = 0,
	CUR = 1,
	END = 2,
};

export fn exit(code: i32) void = {
	syscall1(nr.EXIT, code: i64);
};

// PATH_MAX / pathbuf / kpath — port of Hare's ref/hare/sys/+linux/
// syscalls.ha:25,27,29-55. Hare's `path` accepts a sum `(str |
// []u8 | *const u8)`; ww's lib/os public surface narrows to `str`
// (the Hare-faithful surface at ref/hare/os/os.ha:37,47,50 etc).
// Internally, [[kpath]] copies the `str` bytes into a single
// module-level [[pathbuf]] scratch slot and NUL-terminates so the
// raw Linux syscalls (which require C strings) see a valid
// terminator. Same precedent as Hare's static `pathbuf`.
//
// Non-reentrant: one buffer, every [[stat]] / [[open]] / etc.
// rewrites it. Same caveat as strconv's `*tos` family (overwritten
// on next call). Caller must NOT hold a kpath-returned pointer
// across another lib/os path call. Graduates when ww grows a
// thread story.
//
// `nil`-as-overflow over `(*u8 | oserror)`: wwstage over-allocates
// 1-word-payload tagged returns to 24B (cstage emits 16B).
// Task #9; revert at task #10 when fixed. Repro at
// .ai/probe_tagged_return_pointer_payload.ww.
export def PATH_MAX: i32 = 4096;
let pathbuf: [4096]u8;

fn kpath(p: str) *u8 = {
	if (p.len + 1 >= PATH_MAX) { return nil: *u8; };	// ENAMETOOLONG
	let i: i32 = 0;
	for (i < p.len) { pathbuf[i] = p[i]; i += 1; };
	pathbuf[p.len] = 0u8;
	return &pathbuf[0];
};

// Raw, non-fallible primitives. These return Linux's int conventions
// (negative = -errno, non-negative = bytes/fd/etc). Callers wanting a
// Hare-style fallible API use the wrappers below.
export fn write(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.WRITE, fd: i64, buf: i64, n: i64);
};

export fn read(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.READ, fd: i64, buf: i64, n: i64);
};

export fn close(fd: i32) i32 = {
	return syscall1(nr.CLOSE, fd: i64): i32;
};

// dup2(2): make `newfd` refer to the same description as `oldfd`,
// closing `newfd` first if open. Returns `newfd` on success or a
// negative errno. Used by w6c_ww to redirect stdout into an output
// file without changing the cgen emit path.
export fn dup2(oldfd: i32, newfd: i32) i32 = {
	return syscall2(nr.DUP2, oldfd: i64, newfd: i64): i32;
};

// Fallible wrappers. The error variant is `oserror` (an i64 carrying
// -errno). The sum type makes success/failure explicit and lets
// callers `?` the result up the stack.
export fn tryread(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let r: i64 = read(fd, buf, n);
	if (r < 0) { return r: oserror; };
	return r;
};

export fn trywrite(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let r: i64 = write(fd, buf, n);
	if (r < 0) { return r: oserror; };
	return r;
};

// open — Linux open(2). Returns -errno on failure, fd otherwise.
// Higher-level callers prefer `tryopen`. Mirrors Hare's os::open
// (ref/hare/os/os.ha:117); kpath lands the bytes in pathbuf.
// Returns -ENAMETOOLONG (-36) if the path overflows PATH_MAX.
export fn open(path: str, flags: flag, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };	// ENAMETOOLONG
	return syscall3(nr.OPEN, p: i64, (flags as i32): i64, mode: i64): i32;
};

export fn tryopen(path: str, flags: flag, mode: i32) (i32 | oserror) = {
	let fd: i32 = open(path, flags, mode);
	if (fd < 0) { return fd: i64: oserror; };
	return fd;
};

// lseek — set/inspect the fd's position. Returns the new offset or
// a negative errno. We use this for fstat-free file-size discovery
// (open ⇒ lseek to end ⇒ lseek back).
export fn lseek(fd: i32, off: i64, w: whence) i64 = {
	return syscall3(nr.LSEEK, fd: i64, off, (w as i32): i64);
};

// oserror — the underlying errno from a failed syscall, as a
// negative i64 (Linux's int convention; e.g. -2 = ENOENT). The
// `!`-flagged alias makes ?-propagation pick this variant as the
// error half of any (T | oserror) shape. Hare's analogue is
// errors::errno carried inside io::error.
export type oserror = !i64;

// filesize — byte length of an open fd via lseek-to-end-and-back.
export fn filesize(fd: i32) (i64 | oserror) = {
	let end: i64 = lseek(fd, 0i64, whence.END);
	if (end < 0) { return end: oserror; };
	let r: i64 = lseek(fd, 0i64, whence.SET);
	if (r < 0) { return r: oserror; };
	return end;
};

// readall — keep reading until `n` bytes have arrived or the fd
// closes early. Hare name (io::readall); the buffer is caller-
// supplied, matching the Plan 9 subset convention.
export fn readall(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let got: u64 = 0u64;
	for (got < n) {
		let r: i64 = read(fd, buf + got, n - got);
		if (r < 0) { return r: oserror; };
		if (r == 0) { return got: i64; };	// short read: caller decides
		got += r: u64;
	};
	return got: i64;
};

// writeall — keep writing until `n` bytes have been accepted or the
// fd refuses progress. Hare name (io::writeall).
export fn writeall(fd: i32, buf: *u8, n: u64) (i64 | oserror) = {
	let sent: u64 = 0u64;
	for (sent < n) {
		let r: i64 = write(fd, buf + sent, n - sent);
		if (r < 0) { return r: oserror; };
		if (r == 0) { return sent: i64; };
		sent += r: u64;
	};
	return sent: i64;
};

// ---- process and filesystem helpers used by the `ww` driver ----------

// access(2): returns 0 if the file is reachable, negative errno
// otherwise. mode is the bitset described in <unistd.h> (F_OK=0).
// Mirrors Hare's os::access (ref/hare/os/+linux/fs.ha:access).
// Returns -ENAMETOOLONG (-36) if the path overflows PATH_MAX.
export fn access(path: str, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall2(nr.ACCESS, p: i64, mode: i64): i32;
};

// remove — unlink(2). Mirrors Hare's os::remove
// (ref/hare/os/os.ha:12).
export fn remove(path: str) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall1(nr.UNLINK, p: i64): i32;
};

// mkdir — mkdir(2). Mode is the unix permission bitset (e.g. 0o700).
// Returns 0 on success, negative errno otherwise. Mirrors Hare's
// os::mkdir (ref/hare/os/os.ha:50).
export fn mkdir(path: str, mode: i32) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall2(nr.MKDIR, p: i64, mode: i64): i32;
};

// rmdir — rmdir(2). Mirrors Hare's os::rmdir
// (ref/hare/os/os.ha:58).
export fn rmdir(path: str) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall1(nr.RMDIR, p: i64): i32;
};

// mkdirs — recursive mkdir. Creates `path` and any non-existent
// parent directories with the given mode. EEXIST is silently
// accepted (matches Hare's `errors::exists` skip in os::mkdirs);
// any other syscall failure surfaces as `oserror`.
//
// Mirrors Hare's os::mkdirs (ref/hare/os/os.ha:54). The in-place
// '/' → NUL splice walks the kpath-loaded [[pathbuf]] directly
// instead of recursing through [[mkdir]] — re-entering kpath would
// clobber the buffer mid-walk (single static slot, see kpath's
// non-reentrancy note above).
export fn mkdirs(path: str, mode: i32) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let n: i32 = path.len;
	if (n == 0) { return; };

	// Walk forward; at each '/' boundary, NUL-terminate the prefix,
	// raw MKDIR syscall on pathbuf, restore the slash, continue.
	// Skip index 0 so a leading '/' on absolute paths doesn't
	// trigger an empty mkdir.
	let i: i32 = 1;
	for (i < n) {
		if (pathbuf[i] == 47u8) {	// '/'
			pathbuf[i] = 0u8;
			let r: i32 = syscall2(nr.MKDIR,
			    (&pathbuf[0]): i64, mode: i64): i32;
			pathbuf[i] = 47u8;
			if (r < 0) {
				if (r != -17) { return r: i64: oserror; };
			};
		};
		i += 1;
	};

	let r: i32 = syscall2(nr.MKDIR,
	    (&pathbuf[0]): i64, mode: i64): i32;
	if (r < 0) {
		if (r != -17) { return r: i64: oserror; };
	};
	return;
};

// getpid(2). Used by the driver to mint unique scratch paths.
export fn getpid() i32 = {
	return syscall0(nr.GETPID): i32;
};

// fork(2): 0 in the child, child pid in the parent, negative errno
// on failure.
export fn fork() i32 = {
	return syscall0(nr.FORK): i32;
};

// execve(2): on success, does not return. Mirrors Hare's
// os::exec::exec path arg (str). argv/envp stay `**u8` — the
// kernel takes a NUL-pointer-terminated table of NUL-terminated
// C strings, a different shape from a path.
export fn execve(path: str, argv: **u8, envp: **u8) i32 = {
	let p: *u8 = kpath(path);
	if (p == nil: *u8) { return -36i32; };
	return syscall3(nr.EXECVE, p: i64, argv: i64, envp: i64): i32;
};

// wait4(2): wait for `pid` (or any child if -1), store status in
// `*status`, return the pid that ended (or negative errno).
export fn wait4(pid: i32, status: *i32, options: i32, rusage: *void) i32 = {
	return syscall4(nr.WAIT4, pid: i64, status: i64,
	    options: i64, rusage: i64): i32;
};

// getcwd(2) — Linux flavour. Writes the NUL-terminated cwd into `buf`
// and returns the number of bytes written (including the NUL), or a
// negative errno. The driver uses it to expand `.` to the cwd's
// basename for `ww build` / `ww test`.
export fn getcwd(buf: *u8, n: u64) i64 = {
	return syscall2(nr.GETCWD, buf: i64, n: i64);
};

// getdents64(2) — Linux directory enumeration. The fd must be opened
// with O_RDONLY on a directory. `buf` receives a packed sequence of
// linux_dirent64 records:
//
//   struct linux_dirent64 {
//       u64  d_ino;       // 0..7
//       i64  d_off;       // 8..15
//       u16  d_reclen;    // 16..17  — total bytes for this record
//       u8   d_type;      // 18      — DT_REG/DT_DIR/...
//       u8   d_name[];    // 19..    — NUL-terminated name + padding
//   };
//
// Returns bytes written into `buf` (advance by d_reclen to walk),
// 0 at end-of-directory, or a negative errno.
export fn getdents64(fd: i32, buf: *u8, n: u64) i64 = {
	return syscall3(nr.GETDENTS64, fd: i64, buf: i64, n: i64);
};

// ---- environment ------------------------------------------------------

// rt_envp — runtime-side getter. rt/start.s captures envp into a DATAW
// slot before calling main; this binding lifts the captured pointer
// into ww. Same FFI shape as rt_syscall / rt_alloc / rt_abort: a TEXT
// symbol the linker resolves. The returned `**u8` is a NUL-terminated
// table of `*u8` entries, each pointing at a NUL-terminated
// "NAME=VALUE" byte sequence.
//
// We don't expose `rtenvp` directly; [[getenv]] is the only consumer.
@symbol("rt_envp") fn rtenvp() **u8;

// getenv — POSIX getenv. Returns a borrowed `str` view over the value
// bytes of the named environment variable, or void if the name is not
// present. The view is valid for the process lifetime — the bytes
// live in the kernel-supplied envp table at process entry. A future
// `setenv` (separate task) that grows the table behind the scenes
// would invalidate prior views; v1 has no setenv, so callers can
// hold the view indefinitely.
//
// Mirrors Hare's os::tryenv shape (returns void rather than panicking
// on missing). Hare also ships os::getenv (`(str | void)`) and
// os::mustenv (panic-on-missing); ww collapses to the single
// `(str | void)` form for now — consumers wanting "must" semantics
// abort at the call site.
//
// Algorithm: walk the NUL-pointer-terminated `environ` table doing a
// "name=" prefix match against each entry, byte-wise. NUL inside
// `name` would never match a real env var (env var names cannot
// contain '\0'), so we don't filter — POSIX puts that responsibility
// on the caller.
export fn getenv(name: str) (str | void) = {
	let envp: **u8 = rtenvp();
	let i: i32 = 0;
	for (true) {
		let entry: *u8 = envp[i];
		if (entry == nil: *u8) { return; };
		let j: i32 = 0;
		let matched: bool = true;
		for (j < name.len) {
			if (entry[j] == 0u8) { matched = false; break; };
			if (entry[j] != name[j]) { matched = false; break; };
			j += 1;
		};
		if (matched) {
			if (entry[name.len] == 61u8) {	// '='
				let val: *u8 = entry + ((name.len + 1): u64);
				let n: i32 = 0;
				for (val[n] != 0u8) { n += 1; };
				let r: str;
				r.ptr = val;
				r.len = n;
				return r;
			};
		};
		i += 1;
	};
	return;
};

// ---- stat / lstat / fstat / exists -----------------------------------
//
// Ports of Hare's stat family (ref/hare/fs/fs.ha:172,196 +
// ref/hare/sys/+linux/stat.ha:24-58). The Hare surface returns
// `filestat` by value; ww's cgreturn ABI tops out at 24B today (see
// STATUS task #21) and filestat is 80B, so [[stat]] / [[lstat]] /
// [[fstat]] take an out-parameter and return `(void | oserror)`.
// Re-evaluate the by-value shape when full sret lands.
//
// `filestat`, `mode`, and `stat_mask` live in lib/os because ww has
// no lib/fs yet; Hare puts them in `fs::`. These types graduate to
// lib/fs when that module ships — callers should expect a future
// re-export.
//
// Underlying syscall is SYS_newfstatat (262), which unifies
// stat/lstat/fstat through the `dirfd + flags` triple:
//   stat  = newfstatat(AT_FDCWD, path,  0)
//   lstat = newfstatat(AT_FDCWD, path,  AT_SYMLINK_NOFOLLOW)
//   fstat = newfstatat(fd,       "",    AT_EMPTY_PATH)
// Avoiding SYS_statx — its 256B variable layout would buy btime,
// but Hare's filestat doesn't expose btime either, so we stay on
// the simpler 144B kernel struct.

// fstatat(2) flag values. Linux constants from <linux/fcntl.h>.
// Names mirror Hare's ref/hare/sys/+linux/types.ha:45-51 (capital-
// AT_ prefix, top-level `def`s).
export def AT_FDCWD:            i32 = -100;
export def AT_SYMLINK_NOFOLLOW: i32 = 256;	// 0x100
export def AT_EMPTY_PATH:       i32 = 4096;	// 0x1000

// mode — file-mode bits. Mirrors Hare's fs::mode (ref/hare/fs/
// types.ha:63). Permission bits are the standard Unix octal subset;
// type bits live in the S_IFMT = 0o170000 region. Type-bit test:
//
//   let t: u32 = (fi.mode as u32) & 61440u32;	// 0o170000 mask
//   if (t == os.mode.DIR as u32) { /* directory */ };
//
// Numeric values are octal in Hare's source; ww has no octal
// literals so they're written as decimal with the octal in a
// trailing comment.
export type mode = enum u32 {
	// permission bits
	USER_RWX  = 448u32,	// 0o700
	USER_RW   = 384u32,	// 0o600
	USER_RX   = 320u32,	// 0o500
	USER_R    = 256u32,	// 0o400
	USER_W    = 128u32,	// 0o200
	USER_X    = 64u32,	// 0o100
	GROUP_RWX = 56u32,	// 0o070
	GROUP_RW  = 48u32,	// 0o060
	GROUP_RX  = 40u32,	// 0o050
	GROUP_R   = 32u32,	// 0o040
	GROUP_W   = 16u32,	// 0o020
	GROUP_X   = 8u32,	// 0o010
	OTHER_RWX = 7u32,	// 0o007
	OTHER_RW  = 6u32,	// 0o006
	OTHER_RX  = 5u32,	// 0o005
	OTHER_R   = 4u32,	// 0o004
	OTHER_W   = 2u32,	// 0o002
	OTHER_X   = 1u32,	// 0o001
	SETUID    = 2048u32,	// 0o4000
	SETGID    = 1024u32,	// 0o2000
	STICKY    = 512u32,	// 0o1000
	// file-type bits (S_IFMT mask = 0o170000 = 61440)
	UNKNOWN   = 0u32,
	FIFO      = 4096u32,	// 0o010000
	CHR       = 8192u32,	// 0o020000
	DIR       = 16384u32,	// 0o040000
	BLK       = 24576u32,	// 0o060000
	REG       = 32768u32,	// 0o100000
	LINK      = 40960u32,	// 0o120000
	SOCK      = 49152u32,	// 0o140000
};

// stat_mask — which filestat fields the call populated. Mirrors
// Hare's fs::stat_mask (ref/hare/fs/types.ha:129). newfstatat fills
// every field, so [[stat]] / [[lstat]] / [[fstat]] always set all
// seven bits OR-folded (see [[fillfilestat]]); per-bit testing is
// the documented sparse-backend pattern (cf. Hare's fs::fs network
// backends that only populate mtime+size).
export type stat_mask = enum u32 {
	UID   = 1u32,
	GID   = 2u32,
	SIZE  = 4u32,
	INODE = 8u32,
	ATIME = 16u32,
	MTIME = 32u32,
	CTIME = 64u32,
};

// filestat — Hare's fs::filestat (ref/hare/fs/types.ha:141). 80
// bytes. Times are time.instant (ref/hare/time/instant.ha:9) — the
// canonical Hare shape. See module-header note re: graduation to
// lib/fs.
export type filestat = struct {
	mask:  stat_mask,	// 0  (4)
	mode:  mode,		// 4  (4)
	uid:   u32,		// 8  (4)
	gid:   u32,		// 12 (4)
	sz:    u64,		// 16 (8)
	inode: u64,		// 24 (8)
	atime: time.instant,	// 32 (16)
	mtime: time.instant,	// 48 (16)
	ctime: time.instant,	// 64 (16) — ends at 80
};

// kstat — x86_64 kernel `struct stat` layout. Mirrors
// arch/x86/include/uapi/asm/stat.h (`__kernel_ulong_t`-keyed
// fields). 144 bytes. Module-internal; SYS_newfstatat writes into
// this buffer and the public stat fns then copy the bits into the
// Hare-shaped [[filestat]].
type kstat = struct {
	dev:        u64,	// 0
	ino:        u64,	// 8
	nlink:      u64,	// 16
	mode:       u32,	// 24
	uid:        u32,	// 28
	gid:        u32,	// 32
	pad0:       u32,	// 36
	rdev:       u64,	// 40
	sz:         i64,	// 48
	blksize:    i64,	// 56
	blocks:     i64,	// 64
	atime_sec:  i64,	// 72
	atime_nsec: i64,	// 80
	mtime_sec:  i64,	// 88
	mtime_nsec: i64,	// 96
	ctime_sec:  i64,	// 104
	ctime_nsec: i64,	// 112
	unused0:    i64,	// 120
	unused1:    i64,	// 128
	unused2:    i64,	// 136 — ends at 144
};

// emptypath — single-NUL byte used as the `pathname` arg to
// newfstatat with AT_EMPTY_PATH. The kernel requires a non-NULL
// pointer to a zero-length C string, NOT a null pointer. Bytes are
// read-only from the kernel's view; ww has no module-level const so
// this is a writable `let`.
let emptypath: [1]u8 = [0u8];

// fillfilestat — copy a 144B kstat into the 80B Hare-shaped
// filestat. Internal helper used by all three public entry points.
// Mirrors Hare's st_to_filestat (ref/hare/os/+linux/dirfdfs.ha:259):
// newfstatat populates every field, so the mask is the OR-fold of
// all seven Hare stat_mask bits.
fn fillfilestat(out: *filestat, k: *kstat) void = {
	out.mask        = stat_mask.UID | stat_mask.GID | stat_mask.SIZE
	    | stat_mask.INODE | stat_mask.ATIME | stat_mask.MTIME
	    | stat_mask.CTIME;
	out.mode        = k.mode: mode;
	out.uid         = k.uid;
	out.gid         = k.gid;
	out.sz          = k.sz: u64;
	out.inode       = k.ino;
	out.atime.sec   = k.atime_sec;
	out.atime.nsec  = k.atime_nsec;
	out.mtime.sec   = k.mtime_sec;
	out.mtime.nsec  = k.mtime_nsec;
	out.ctime.sec   = k.ctime_sec;
	out.ctime.nsec  = k.ctime_nsec;
};

// stat — fill *out with metadata for `path`. Follows symlinks.
// Returns ENAMETOOLONG (-36) as `oserror` if the path overflows
// PATH_MAX.
//
// Mirrors Hare's sys::stat (ref/hare/sys/+linux/stat.ha:51) modulo
// the out-param shape forced by the cgreturn 24B cap. Note: Hare's
// higher-level fs::stat (ref/hare/fs/fs.ha:172) instead has lstat
// semantics — we follow sys::stat's POSIX-stat behavior here.
export fn stat(out: *filestat, path: str) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64, 0i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// lstat — like [[stat]] but does NOT follow a terminal symlink.
// Mirrors Hare's sys::lstat (ref/hare/sys/+linux/stat.ha:57).
export fn lstat(out: *filestat, path: str) (void | oserror) = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return -36i64: oserror; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64,
	    AT_SYMLINK_NOFOLLOW: i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// fstat — like [[stat]] but addresses the file by fd. Uses
// newfstatat(fd, "", AT_EMPTY_PATH); the kernel resolves the fd
// directly. Mirrors Hare's sys::fstat (ref/hare/sys/+linux/stat.ha:54).
export fn fstat(out: *filestat, fd: i32) (void | oserror) = {
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    fd: i64, (&emptypath[0]): i64, (&k): i64,
	    AT_EMPTY_PATH: i64);
	if (r < 0) { return r: oserror; };
	fillfilestat(out, &k);
};

// exists — true if `path` resolves to anything (regular file,
// directory, symlink, ...). Stat-shaped (Hare's `fs::exists`,
// ref/hare/fs/fs.ha:196) — no separate syscall. Symlinks are
// followed; a dangling symlink is `false`. ENAMETOOLONG is
// swallowed as `false` — Hare's os::exists doc says "true if a
// node exists at the given path, or false if not."
//
// Race warning: prefer "open and handle the error" over "exists
// then open" in real code (Hare's docstring carries the same
// note). The race is unavoidable in this shape.
//
// Goes through SYS_newfstatat directly rather than match'ing on
// [[stat]]'s `(void | oserror)` return. Functionally identical;
// the direct shape sidesteps a cstage/wwstage cgen disagreement
// on the slot size of `(void | oserror)` (cstage 16B, wwstage 24B
// — same class as STATUS #22, surfaced first time a match on this
// shape combined with an 80B local-struct local frame). Use the
// match shape once #22 lands.
export fn exists(path: str) bool = {
	let cp: *u8 = kpath(path);
	if (cp == nil: *u8) { return false; };
	let k: kstat;
	let r: i64 = syscall4(nr.NEWFSTATAT,
	    AT_FDCWD: i64, cp: i64, (&k): i64, 0i64);
	return r >= 0i64;
};

// types — integer limits. Mirrors Hare's types::limits (I8_MAX, …)
// platform-fixed for amd64. Numeric helpers live in lib/math, matching
// Hare's split between types::limits and math::.

package types;

def I8_MAX:  i8  = 127;
def I16_MAX: i16 = 32767;
def I32_MAX: i32 = 2147483647;
def I64_MAX: i64 = 9223372036854775807;

def I8_MIN:  i8  = -128;
def I16_MIN: i16 = -32768;
def I32_MIN: i32 = -2147483648;
def I64_MIN: i64 = -9223372036854775808;

def U8_MAX:  u8  = 255;
def U16_MAX: u16 = 65535;
def U32_MAX: u32 = 4294967295;
def U64_MAX: u64 = 18446744073709551615;

// bytes — slice operations over []u8. Mirrors Hare's bytes module
// (ref/hare/bytes/) for the in-tree subset: search/equality/prefix
// helpers used by lib/encoding, lib/bufio, lib/memio.
//
// Documented divergences from Hare:
//   - index_slice / rindex_slice use naive O(n·m); Hare specialises
//     2/3/4-byte needles and falls back to two_way (Crochemore-Perrin)
//     for longer (ref/hare/bytes/index.ha:61, ref/hare/bytes/two_way.ha).
//     Correctness equivalent.
//   - peek_token dispatches index/rindex by branching on `reverse`
//     rather than a function-pointer `ifunc` (ref/hare/bytes/tokenize.ha:97).
//     ww has no fn pointers in scope yet — same pattern as lib/strings
//     `move`. Outwardly identical.
//   - tokenize / rtokenize zero the `delim` field on the constructed
//     tokenizer when `in` is empty, rather than mutating the variadic
//     param before the struct write (ref/hare/bytes/tokenize.ha:26-28).
//     Semantically identical; the variadic param is borrowed and
//     captured-by-value into the struct, so mutating either side
//     yields the same observable state.

package bytes;

import os;
import types;

// done — iteration sentinel returned by next_token / peek_token at
// end-of-input. ref/hare/bytes/tokenize.ha uses the built-in `done`
// token; ww spells it per-package the same way lib/encoding/utf8 does
// (utf8.ww:36). Plain `void` (not `!void`): continuation signal.
export type done = void;

// tokenizer — cursor over an input slice. Layout mirrors
// ref/hare/bytes/tokenize.ha:6-10. `p` is the cached peek-position;
// I64_MAX (forward) / I64_MIN (reverse) are the unprimed sentinels.
// p < 0 also identifies a reverse-direction iterator.
export type tokenizer = struct {
	in:    []u8,
	delim: []u8,
	p:     i64,
};

// equal — true iff `a` and `b` have the same length and contents.
// ref/hare/bytes/equal.ha:9.
export fn equal(a: []u8, b: []u8) bool = {
	if (a.len != b.len) { return false; };
	let i: i32 = 0;
	for (i < a.len) {
		if (a[i] != b[i]) { return false; };
		i += 1;
	};
	return true;
};

// index — first offset of `needle` in `s`. u8 needle scans for the
// byte; []u8 needle scans for the substring. void if absent.
// ref/hare/bytes/index.ha:6.
export fn index(s: []u8, needle: (u8 | []u8)) (i32 | void) = {
	match (needle) {
	case let c: u8 => {
		let i: i32 = 0;
		for (i < s.len) {
			if (s[i] == c) { return i; };
			i += 1;
		};
		return;
	};
	case let sub: []u8 => {
		if (sub.len == 0) { return 0; };
		if (sub.len > s.len) { return; };
		let last: i32 = s.len - sub.len;
		let i: i32 = 0;
		for (i <= last) {
			let j: i32 = 0;
			let ok: bool = true;
			for (j < sub.len) {
				if (s[i + j] != sub[j]) { ok = false; j = sub.len; }
				else { j += 1; };
			};
			if (ok) { return i; };
			i += 1;
		};
		return;
	};
	};
	return;
};

// rindex — last offset of `needle` in `s`. Empty []u8 needle returns
// s.len (ref/hare/bytes/index.ha:103 — Hare's loop yields r-0 at i=0).
// ref/hare/bytes/index.ha:86.
export fn rindex(s: []u8, needle: (u8 | []u8)) (i32 | void) = {
	match (needle) {
	case let c: u8 => {
		let i: i32 = s.len - 1;
		for (i >= 0) {
			if (s[i] == c) { return i; };
			i -= 1;
		};
		return;
	};
	case let sub: []u8 => {
		if (sub.len == 0) { return s.len; };
		if (sub.len > s.len) { return; };
		let i: i32 = s.len - sub.len;
		for (i >= 0) {
			let j: i32 = 0;
			let ok: bool = true;
			for (j < sub.len) {
				if (s[i + j] != sub[j]) { ok = false; j = sub.len; }
				else { j += 1; };
			};
			if (ok) { return i; };
			i -= 1;
		};
		return;
	};
	};
	return;
};

// contains — true iff any of `needles` (byte or sub-slice) appears in `s`.
// ref/hare/bytes/contains.ha:6.
export fn contains(s: []u8, needles: (u8 | []u8)...) bool = {
	let i: i32 = 0;
	for (i < needles.len) {
		match (needles[i]) {
		case let b: u8   => {
			match (index(s, b)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		case let n: []u8 => {
			match (index(s, n)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		};
		i += 1;
	};
	return false;
};

// hasprefix — true iff `s` starts with `pre`.
// ref/hare/bytes/contains.ha:21.
export fn hasprefix(s: []u8, pre: []u8) bool = {
	if (pre.len > s.len) { return false; };
	let i: i32 = 0;
	for (i < pre.len) {
		if (s[i] != pre[i]) { return false; };
		i += 1;
	};
	return true;
};

// hassuffix — true iff `s` ends with `suf`.
// ref/hare/bytes/contains.ha:35.
export fn hassuffix(s: []u8, suf: []u8) bool = {
	if (suf.len > s.len) { return false; };
	let off: i32 = s.len - suf.len;
	let i: i32 = 0;
	for (i < suf.len) {
		if (s[off + i] != suf[i]) { return false; };
		i += 1;
	};
	return true;
};

// reverse — in-place reverse of `s`. ref/hare/bytes/reverse.ha:5.
export fn reverse(s: []u8) void = {
	let i: i32 = 0;
	let j: i32 = s.len - 1;
	for (i < j) {
		let t: u8 = s[i];
		s[i] = s[j];
		s[j] = t;
		i += 1;
		j -= 1;
	};
};

// zero — set every byte of `s` to 0. ref/hare/bytes/zero.ha:5.
export fn zero(s: []u8) void = {
	let i: i32 = 0;
	for (i < s.len) {
		s[i] = 0u8;
		i += 1;
	};
};

// tokenize — iterator yielding tokens from `in` separated by any byte
// in `delim`. Leading / trailing / adjacent delims yield empty tokens.
// `delim` is borrowed; caller keeps it valid for the tokenizer's
// lifetime. ref/hare/bytes/tokenize.ha:22.
export fn tokenize(in: []u8, delim: u8...) tokenizer = {
	os.assert(delim.len > 0, "bytes.tokenize called with empty slice");
	os.assert((in.len: i64) < types.I64_MAX,
		"bytes.tokenize: input length exceeds I64_MAX");
	let t: tokenizer;
	t.in = in;
	t.delim = delim;
	if (in.len == 0) {
		t.delim.len = 0;
		t.delim.cap = 0;
	};
	t.p = types.I64_MAX;
	return t;
};

// rtokenize — reverse-direction tokenize. First next_token yields the
// last token, last next_token yields the first. ref/hare/bytes/tokenize.ha:40.
export fn rtokenize(in: []u8, delim: u8...) tokenizer = {
	os.assert(delim.len > 0, "bytes.rtokenize called with empty slice");
	os.assert((in.len: i64) < types.I64_MAX,
		"bytes.rtokenize: input length exceeds I64_MAX");
	let t: tokenizer;
	t.in = in;
	t.delim = delim;
	if (in.len == 0) {
		t.delim.len = 0;
		t.delim.cap = 0;
	};
	t.p = types.I64_MIN;
	return t;
};

// peek_token — next token without advancing the cursor. Returns done
// once `s.delim` has been zeroed by a prior past-end next_token.
// ref/hare/bytes/tokenize.ha:91.
export fn peek_token(s: *tokenizer) ([]u8 | done) = {
	if (s.delim.len == 0) {
		let d: done; return d;
	};

	let reverse: bool = s.p < 0i64;
	let known: bool = false;
	if (reverse) {
		if (s.p != types.I64_MIN) { known = true; };
	} else {
		if (s.p != types.I64_MAX) { known = true; };
	};
	if (!known) {
		let i: i64 = types.I64_MAX;
		if (reverse) { i = types.I64_MIN; };
		let dlen: i64 = 0i64;
		let slen: i64 = s.in.len: i64;

		let k: i32 = 0;
		for (k < s.delim.len) {
			let d: u8 = s.delim[k];
			let ix_found: bool = false;
			let ix_val: i32 = 0;
			if (reverse) {
				match (rindex(s.in, d)) {
				case let v: i32 => { ix_found = true; ix_val = v; };
				case void       => void;
				};
			} else {
				match (index(s.in, d)) {
				case let v: i32 => { ix_found = true; ix_val = v; };
				case void       => void;
				};
			};
			if (ix_found) {
				if (!reverse) {
					if ((ix_val: i64) < i) { i = ix_val: i64; dlen = 1i64; };
				} else {
					if ((ix_val: i64) > i) { i = ix_val: i64; dlen = 1i64; };
				};
			} else {
				if (!reverse) {
					if (slen < i) { i = slen; };
				} else {
					if (0i64 > i) { i = 0i64; };
				};
			};
			k += 1;
		};

		if (reverse) {
			if (i == slen) {
				s.p = -(slen + 1i64);
			} else {
				s.p = i + dlen - slen - 1i64;
			};
		} else {
			s.p = i;
		};
	};

	let r: []u8;
	if (reverse) {
		let start: i32 = (s.in.len: i64 + s.p + 1i64): i32;
		r.ptr = s.in.ptr + (start: u64);
		r.len = s.in.len - start;
		r.cap = r.len;
	} else {
		let end: i32 = s.p: i32;
		r.ptr = s.in.ptr;
		r.len = end;
		r.cap = end;
	};
	return r;
};

// next_token — current token, then advance past it and the delim.
// Once the input is exhausted, returns done and zeros `s.delim` so
// subsequent peeks short-circuit. ref/hare/bytes/tokenize.ha:59.
export fn next_token(s: *tokenizer) ([]u8 | done) = {
	let b: []u8;
	match (peek_token(s)) {
	case let v: []u8 => { b = v; };
	case done        => { let d: done; return d; };
	};

	let slen: i64 = s.in.len: i64;
	let reverse: bool = s.p < 0i64;
	if (reverse) {
		if (slen + s.p + 1i64 == 0i64) {
			s.delim.len = 0;
			s.delim.cap = 0;
			s.in.len = 0;
			s.in.cap = 0;
		} else {
			let end: i32 = (slen + s.p + 1i64 - 1i64): i32;
			s.in.len = end;
			s.in.cap = end;
		};
		s.p = types.I64_MIN;
	} else {
		if (s.p == slen) {
			s.delim.len = 0;
			s.delim.cap = 0;
			s.in.len = 0;
			s.in.cap = 0;
		} else {
			let adv: u64 = (s.p: u64) + 1u64;
			let adv_i32: i32 = (s.p: i32) + 1;
			s.in.ptr = s.in.ptr + adv;
			s.in.len = s.in.len - adv_i32;
			s.in.cap = s.in.cap - adv_i32;
		};
		s.p = types.I64_MAX;
	};
	return b;
};

// remaining_tokens — the unconsumed portion of `s.in`. Read-only view.
// ref/hare/bytes/tokenize.ha:145.
export fn remaining_tokens(s: *tokenizer) []u8 = {
	return s.in;
};

// rt_ensure is the runtime slice-growth helper invoked by the
// `append(s, v)` builtin. We bind it directly because the builtin's
// expansion stores only 8 bytes of the new element (cgen emits a
// single MOVQ), losing the .len/.cap fields of a []u8 element (24B).
// Mirrors the same workaround in lib/shlex.shlex (appendstr, 16B) and
// lib/getopt.getopt (appendoption, 24B); collapses in one go when the
// append builtin learns to store the full element width.
@symbol("rt_ensure") fn rtensure(s: *void, membsz: u64) void;

// appendslice — grow `*slice` by one and store `item` (24B). Mirror
// of [[shlex.appendstr]] / [[getopt.appendoption]]. Bypasses the
// `append` builtin's first-8B-only-store gap for a slice-element.
fn appendslice(slice: *[][]u8, item: []u8) void = {
	let newlen: i32 = slice.len + 1;
	slice.len = newlen;
	rtensure(slice: *void, 24u64);
	let dst: *[]u8 = &slice.ptr[newlen - 1];
	dst.ptr = item.ptr;
	dst.len = item.len;
	dst.cap = item.cap;
};

// splitn — split `in` on any byte in `delim`, returning up to `n`
// tokens via forward iteration. The trailing slot (when more than
// `n - 1` tokens exist) holds the unconsumed remainder.
//
// The caller frees the returned slice via
// `os.free(r.ptr: *void, (r.cap: u64) * 24u64)`. Element bytes are
// borrowed from `in`.
//
// Hare's `([][]u8 | nomem)` collapses to `[][]u8` here: ww os.alloc
// has no recoverable failure path. Same precedent as
// shlex.split / getopt.tryparse.
//
// ref/hare/bytes/tokenize.ha:156.
export fn splitn(in: []u8, delim: []u8, n: i32) [][]u8 = {
	os.assert(delim.len > 0,
		"bytes.splitn must not be called with an empty delimiter");
	let toks: [][]u8;
	toks.ptr = nil: *[]u8;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = tokenize(in, delim...);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: []u8 => { appendslice(&toks, s); };
		case done        => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case done         => void;
	case let pk: []u8 => {
		let r: []u8 = remaining_tokens(&tok);
		appendslice(&toks, r);
	};
	};
	return toks;
};

// rsplitn — reverse-direction counterpart to [[splitn]]: tokens are
// collected from the end of `in`. The trailing slot holds the
// unconsumed prefix (everything before the n-th-from-last delim hit).
//
// When the input has fewer than n tokens, the `done` short-circuit
// returns toks UN-reversed (in last-token-first order). Mirrors Hare
// at ref/hare/bytes/tokenize.ha:196-199 where the in-place reverse
// step is gated behind the n-1 loop running to completion. Only the
// "loop ran to completion AND peek saw a remainder" path applies the
// reverse; both early-exit paths skip it.
//
// ref/hare/bytes/tokenize.ha:186.
export fn rsplitn(in: []u8, delim: []u8, n: i32) [][]u8 = {
	os.assert(delim.len > 0,
		"bytes.rsplitn called with empty delimiter");
	let toks: [][]u8;
	toks.ptr = nil: *[]u8;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = rtokenize(in, delim...);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: []u8 => { appendslice(&toks, s); };
		case done        => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case done         => void;
	case let pk: []u8 => {
		let r: []u8 = remaining_tokens(&tok);
		appendslice(&toks, r);
	};
	};

	// In-place reverse so callers see argv-order, matching Hare
	// (ref/hare/bytes/tokenize.ha:207). Element copy is field-wise
	// through `*[]u8` because `toks[i] = toks[j]` (full 24B slice
	// store) lands in the multi-word-store gap noted at
	// cmd/w6c/cgen.c:6515-6523.
	let a: i32 = 0;
	let b: i32 = toks.len - 1;
	for (a < b) {
		let pa: *[]u8 = &toks.ptr[a];
		let pb: *[]u8 = &toks.ptr[b];
		let tp: *u8 = pa.ptr;
		let tl: i32 = pa.len;
		let tc: i32 = pa.cap;
		pa.ptr = pb.ptr;
		pa.len = pb.len;
		pa.cap = pb.cap;
		pb.ptr = tp;
		pb.len = tl;
		pb.cap = tc;
		a += 1;
		b -= 1;
	};
	return toks;
};

// split — full split of `in` on `delim` (no token cap). Mirrors
// `splitn(in, delim, types::SIZE_MAX)`. ww uses `types.I32_MAX`
// because the index type is i32 (lib/CLAUDE.md).
//
// ref/hare/bytes/tokenize.ha:225.
export fn split(in: []u8, delim: []u8) [][]u8 = {
	return splitn(in, delim, types.I32_MAX);
};

// encoding/utf8 — UTF-8 encode/decode. Hare port; see
// ref/hare/encoding/utf8/{types,rune,encode,decode,decodetable}.ha.
//
// The decoder is Hoehrmann's branchless DFA, originally published
// at <https://bjoern.hoehrmann.de/utf-8/decoder/dfa/>. Hare's
// ref/hare/encoding/utf8/decodetable.ha:4 restructures Hoehrmann's
// flat table to 2D `[8][256]i8`; we flatten back to 1D `[2048]i8`
// because ww cgen does not yet ship 2D arrays (task #20).
//
// Surface deviation from ref/hare/encoding/utf8:
//
//   - `encoderune` takes a caller-supplied `out: []u8` and returns
//     the byte count. Hare returns a slice into a `static let buf`;
//     the caller-buffer form mirrors lib/encoding/hex.encode and
//     skips the static-buffer/slice-return pair.
//
// Deferred (no in-tree caller, follow-up tasks): `appendrune`,
// `strencode`, `strdecode`. Hare's string-iteration surface
// (`strings::iterator`/`strings::next` — ref/hare/strings/iter.ha)
// lives under lib/strings, not here.

// ref/hare/encoding/utf8/types.ha:6 — incomplete trailing sequence.
// Plain `void` (not `!void`): a truncated tail is a control-flow
// signal, not an error caller can ignore.
package utf8;

export type more = void;

// ref/hare/encoding/utf8/types.ha:9 — invalid UTF-8 sequence.
export type invalid = !void;

// `done` is not a built-in singleton in ww (Hare ships it as part of
// the type system). Plain `void` (not `!void`): end-of-input is a
// continuation signal, not an error. lib/io spells its EOF the same
// way (lib/io/io.ww:8-11).
export type done = void;

// ref/hare/encoding/utf8/decodetable.ha:4 — Hoehrmann's UTF-8 DFA,
// flat 1D `[2048]i8`. Layout: dfa[state*256 + byte] gives the next
// state (>0), the accept transition (0 — emit rune), or invalid (-1).
// Values match ref/hare/encoding/utf8/decodetable.ha verbatim.
let dfa: [2048]i8 = [
	// state 0 — initial byte: ASCII accepts (0), continuation/illegal
	// byte rejects (-1), legal multibyte start emits a state.
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	3i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 4i8, 2i8, 2i8,
	5i8, 6i8, 6i8, 6i8, 7i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 1 — expecting one continuation byte (0x80..0xBF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8, 0i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 2 — expecting one continuation byte (full 0x80..0xBF range).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 3 — first byte was 0xE0; continuation byte must be 0xA0..0xBF
	// (rejects overlong 3-byte encodings).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 4 — first byte was 0xED; continuation byte must be 0x80..0x9F
	// (rejects UTF-16 surrogate codepoints U+D800..U+DFFF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8, 1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 5 — first byte was 0xF0; continuation byte must be 0x90..0xBF
	// (rejects overlong 4-byte encodings).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 6 — middle continuation byte of a 4-byte sequence (0x80..0xBF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,

	// state 7 — first byte was 0xF4; continuation byte must be 0x80..0x8F
	// (rejects codepoints above U+10FFFF).
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8, 2i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
	-1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8, -1i8,
];

// ref/hare/encoding/utf8/decode.ha:17 — payload-bit masks. Hare's
// [2][8]u8 flattened to 1D [16]u8; row 0 (offsets 0..7) is the
// continuation-byte mask (always 0x3F), row 1 (offsets 8..15) is the
// initial-byte payload mask indexed by the transition class.
let masks: [16]u8 = [
	0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8, 0x3fu8,
	0x7fu8, 0x1fu8, 0x0fu8, 0x0fu8, 0x0fu8, 0x07u8, 0x07u8, 0x07u8,
];

// ref/hare/encoding/utf8/decode.ha:6 — incremental decoder state.
export type decoder = struct {
	offs: i32,
	src:  []u8,
};

// ref/hare/encoding/utf8/decode.ha:12.
export fn decode(src: []u8) decoder = {
	let d: decoder;
	d.src = src;
	d.offs = 0;
	return d;
};

// ref/hare/encoding/utf8/decode.ha:27. Returns the next rune from a
// decoder, `done` at end-of-input, `more` on truncated trailing
// sequence, `invalid` on malformed input (overlong, surrogate,
// out-of-range, bad continuation).
//
// Algorithm is verbatim Hoehrmann (see file header). One structural
// rewrite: Hare encodes the "initial vs continuation byte" decision
// as the branchless `(state - 1): uint >> 31`, which assumes a 32-bit
// uint. ww's uint is 64-bit (cmd/wcc/type.c:58), so the shift answer
// would be 0x1_ffff_ffff rather than 1. We spell the same predicate
// with an explicit conditional.
export fn next(d: *decoder) (rune | done | more | invalid) = {
	if (d.offs == d.src.len) {
		let dn: done; return dn;
	};
	let nx: i32 = 0;
	let state: i32 = 0;
	let r: u32 = 0u32;
	for (d.offs < d.src.len) {
		let b: u8 = d.src[d.offs];
		let bi: i32 = b: i32;
		let row: i32 = state * 256 + bi;
		let cell: i8 = dfa[row];
		nx = cell: i32;
		let mi: i32 = 0;
		if (state == 0) { mi = 1; };
		let m: u8 = masks[mi * 8 + (nx & 7)];
		r = (r << 6u32) | ((b & m): u32);
		if (nx <= 0) {
			d.offs += 1;
			if (nx == 0) { return r: rune; };
			let e: invalid; return e;
		};
		state = nx;
		d.offs += 1;
	};
	let mr: more; return mr;
};

// ref/hare/encoding/utf8/decode.ha:207. Strict whole-input check.
// The hot path: tight DFA loop, no rune assembly. Bails the moment
// the table returns -1 so malformed inputs don't pay for the rest
// of the buffer.
export fn validate(src: []u8) (void | invalid) = {
	let state: i32 = 0;
	let i: i32 = 0;
	for (i < src.len) {
		if (state < 0) { break; };
		let bi: i32 = src[i]: i32;
		let cell: i8 = dfa[state * 256 + bi];
		state = cell: i32;
		i += 1;
	};
	if (state == 0) { return; };
	let e: invalid; return e;
};

// ref/hare/encoding/utf8/rune.ha:5. Encoded byte length of `r` as
// UTF-8. Callers in ww use this to size the buffer they hand to
// [[encoderune]]; values >0x10FFFF or negative are not legal Unicode
// codepoints and Hare aborts on them in `encoderune` itself, so we
// keep `runesz` infallible (matches Hare).
export fn runesz(r: rune) i32 = {
	let ch: u32 = r: u32;
	if (ch < 128u32)    { return 1; };
	if (ch < 2048u32)   { return 2; };
	if (ch < 65536u32)  { return 3; };
	return 4;
};

// ref/hare/encoding/utf8/rune.ha:15. Expected byte length of the
// codepoint that starts with `c`, or `invalid` if `c` cannot start
// a legal UTF-8 sequence. Constants written in decimal because ww
// doesn't accept Hare's `0b1000_0000` binary syntax: 0x80=128,
// 0xC2=194, 0xE0=224, 0xF0=240, 0xF8=248.
export fn utf8sz(c: u8) (i32 | invalid) = {
	if (c < 128u8) { return 1; };
	if (c < 194u8) { let e: invalid; return e; };
	if (c >= 248u8) { let e: invalid; return e; };
	if (c < 224u8) { return 2; };
	if (c < 240u8) { return 3; };
	return 4;
};

// ref/hare/encoding/utf8/encode.ha:7. Encode `r` into `out` (caller-
// supplied; must hold at least [[runesz]](r) bytes) and return the
// byte count. ABORT if `r` is a UTF-16 surrogate or above U+10FFFF —
// same precondition Hare asserts at ref/hare/encoding/utf8/encode.ha:9.
//
// Surface deviation: Hare returns `[]u8` (slice into a static buf).
// ww uses the caller-buffer form (matches lib/encoding/hex.encode);
// caller can reuse a [4]u8 stack scratch across encodes.
export fn encoderune(out: []u8, r: rune) i32 = {
	let ch: u32 = r: u32;
	if (ch >= 0xD800u32) {
		if (ch <= 0xDFFFu32) {
			abort("utf8.encoderune: surrogate codepoint");
		};
	};
	if (ch > 0x10FFFFu32) {
		abort("utf8.encoderune: codepoint > U+10FFFF");
	};

	let n: i32 = 0;
	let first: u8 = 0u8;
	if (ch < 0x80u32) {
		first = 0u8; n = 1;
	} else if (ch < 0x800u32) {
		first = 0xC0u8; n = 2;
	} else if (ch < 0x10000u32) {
		first = 0xE0u8; n = 3;
	} else {
		first = 0xF0u8; n = 4;
	};

	let v: u32 = ch;
	let i: i32 = n - 1;
	for (i > 0) {
		out[i] = ((v: u8) & 0x3Fu8) | 0x80u8;
		v = v >> 6u32;
		i -= 1;
	};
	out[0] = (v: u8) | first;
	return n;
};

// ref/hare/encoding/utf8/decode.ha:52. Walks back from `d.offs` to a
// byte that could start a codepoint (state-0 dfa cell != -1), re-decodes
// forward from there, and confirms the forward decode lands back at the
// original offset. Returns `done` at start-of-input; `invalid` if no
// initial byte appears within 4 steps (no legal UTF-8 codepoint exceeds
// 4 bytes), if the forward decode returns `more`/`invalid`, or if it
// lands at a different offset than expected. Returns `more` when the
// walk reaches byte 0 without finding any initial byte.
//
// Hare's `for (d.offs < len(d.src); d.offs -= 1)` relies on size_t
// wrap-around to exit when offs underflows past 0; ww's offs is i32,
// so we spell the same exit as `d.offs >= 0`. Hare's `defer d.offs = t`
// is inlined in each match arm — ww has no defer.
export fn prev(d: *decoder) (rune | done | more | invalid) = {
	if (d.offs == 0) {
		let dn: done; return dn;
	};
	let n: i32 = d.offs;
	d.offs -= 1;
	for (d.offs >= 0) {
		let b: u8 = d.src[d.offs];
		let bi: i32 = b: i32;
		let cell: i8 = dfa[bi];
		if (cell: i32 != -1) {
			let t: i32 = d.offs;
			match (next(d)) {
			case let r: rune => {
				let landed: i32 = d.offs;
				d.offs = t;
				if (landed != n) {
					let e: invalid; return e;
				};
				return r;
			};
			case let dn: done => {
				d.offs = t;
				let e: invalid; return e;
			};
			case let m: more => {
				d.offs = t;
				let e: invalid; return e;
			};
			case let e: invalid => {
				d.offs = t;
				let e2: invalid; return e2;
			};
			};
		};
		if (n - d.offs == 4) {
			let e: invalid; return e;
		};
		d.offs -= 1;
	};
	let mr: more; return mr;
};

// ref/hare/encoding/utf8/decode.ha:74. Borrowed view of the bytes from
// the decoder's current position to the end of its source.
export fn remaining(d: *decoder) []u8 = {
	let r: []u8;
	r.ptr = d.src.ptr + (d.offs: u64);
	r.len = d.src.len - d.offs;
	r.cap = d.src.len - d.offs;
	return r;
};

// ref/hare/encoding/utf8/decode.ha:80. Borrowed view of the bytes
// between two decoders' positions. Precondition (Hare asserts both):
// the decoders share the same source, and `begin.offs <= end.offs`.
export fn slice(begin: *decoder, end: *decoder) []u8 = {
	if (begin.src.ptr != end.src.ptr) {
		abort("utf8.slice: decoders from different sources");
	};
	if (begin.offs > end.offs) {
		abort("utf8.slice: begin past end");
	};
	let r: []u8;
	r.ptr = begin.src.ptr + (begin.offs: u64);
	r.len = end.offs - begin.offs;
	r.cap = end.offs - begin.offs;
	return r;
};

// ref/hare/encoding/utf8/decode.ha:203. Byte position of the decoder
// in its source.
export fn position(d: *decoder) i32 = {
	return d.offs;
};


// strings — operations over str ({ptr,len}). Hare port; see
// ref/hare/strings/.
//
// Documented divergences from Hare:
//
//   - `trim` / `ltrim` / `rtrim` 0-arg returns the input unchanged.
//     Hare strips ASCII whitespace via `bytes::ltrim(input,
//     whitespace...)`; that needs `lib/bytes` variadic graduation
//     (future commit).
//   - `byteindex` / `rbyteindex` rune arms encode via
//     `utf8.encoderune`; the legacy impls scanned for `r: u8` (an
//     undocumented ASCII-only restriction that silently dropped
//     to the wrong byte for U+80..U+7FF and higher).
//   - `dup(s: str) str` — Hare returns `(str | nomem)`. ww's
//     `os.alloc` aborts on OOM (no `nomem` type), so we return plain
//     `str`. Empty input returns `{nil, 0}`; Hare returns the static
//     empty string — same observable result.
//   - `iterator` is flattened (`offs`, `src`, `reverse` fields).
//     Hare uses anonymous-embedded `utf8::decoder`
//     (ref/hare/strings/iter.ha:6-9); ww has no anonymous-embed
//     syntax, so `next`/`prev`/`slice` copy `offs`/`src` into a
//     local `utf8.decoder` for the call (and `next`/`prev` write
//     `offs` back).
//   - Hare's private `move()` helper dispatches on a `forward: bool`
//     using a function-pointer `let fun = if (forward) &utf8::next
//     else &utf8::prev`. ww has no fn-pointers in scope yet, so the
//     dispatch is a branch on `forward` selecting the call site.

package strings;

import bytes;
import encoding.utf8;
import os;
import types;

// toutf8 — borrowed []u8 view of `s`. ref/hare/strings/utf8.ha:29.
// `cap` equals `len`; the slice does not own a separate allocation.
export fn toutf8(s: str) []u8 = {
	let r: []u8;
	r.ptr = s.ptr;
	r.len = s.len;
	r.cap = s.len;
	return r;
};

// fromutf8_unsafe — borrowed str view of `in`. Does not validate.
// ref/hare/strings/utf8.ha:10.
export fn fromutf8_unsafe(in: []u8) str = {
	let r: str;
	r.ptr = in.ptr;
	r.len = in.len;
	return r;
};

// compare — three-way bytewise codepoint-order comparison.
// ref/hare/strings/compare.ha:12.
export fn compare(a: str, b: str) i32 = {
	let n: i32 = a.len;
	if (b.len < n) { n = b.len; };
	let i: i32 = 0;
	for (i < n) {
		if (a[i] != b[i]) { return (a[i]: i32) - (b[i]: i32); };
		i += 1;
	};
	return a.len - b.len;
};

// dup — allocate a fresh copy of `s`. Caller releases with
// `os.free(r.ptr, r.len: u64)`. ref/hare/strings/dup.ha:7.
export fn dup(s: str) str = {
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (s.len == 0) { return r; };
	let buf: *u8 = os.alloc(s.len: u64): *u8;
	let i: i32 = 0;
	for (i < s.len) { buf[i] = s[i]; i += 1; };
	r.ptr = buf;
	r.len = s.len;
	return r;
};

// freeall — release each element + the slice header. The natural
// disposer for any `[]str` of dup'd elements (e.g. shlex.split).
// ref/hare/strings/dup.ha:38.
//
// Empty elements (`{nil, 0}` from a zero-length dup) are skipped:
// os.free on a nil pointer at len 0 tickles the rt_free guard. The
// slice header itself is freed at `cap * 16` (one str = 16B); a
// never-grown slice (cap == 0) skips the header free.
export fn freeall(s: []str) void = {
	let i: i32 = 0;
	for (i < s.len) {
		if (s[i].len > 0) {
			os.free(s[i].ptr: *void, s[i].len: u64);
		};
		i += 1;
	};
	if (s.cap > 0) {
		os.free(s.ptr: *void, (s.cap: u64) * 16u64);
	};
};

// concat — fresh allocation containing each element of `strs` in
// order. Caller releases with `os.free(r.ptr, r.len: u64)`.
// ref/hare/strings/concat.ha:5. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM.
export fn concat(strs: str...) str = {
	let total: i32 = 0;
	let i: i32 = 0;
	for (i < strs.len) { total += strs[i].len; i += 1; };
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (total == 0) { return r; };
	let buf: *u8 = os.alloc(total: u64): *u8;
	let off: i32 = 0;
	i = 0;
	for (i < strs.len) {
		let j: i32 = 0;
		for (j < strs[i].len) {
			buf[off + j] = strs[i][j];
			j += 1;
		};
		off += strs[i].len;
		i += 1;
	};
	r.ptr = buf;
	r.len = total;
	return r;
};

// join — fresh allocation with `delim` placed between each element of
// `strs`. Caller releases with `os.free(r.ptr, r.len: u64)`.
// ref/hare/strings/concat.ha:46. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM.
export fn join(delim: str, strs: str...) str = {
	let total: i32 = 0;
	let i: i32 = 0;
	for (i < strs.len) {
		total += strs[i].len;
		if (i + 1 < strs.len) { total += delim.len; };
		i += 1;
	};
	let r: str;
	r.ptr = nil;
	r.len = 0;
	if (total == 0) { return r; };
	let buf: *u8 = os.alloc(total: u64): *u8;
	let off: i32 = 0;
	i = 0;
	for (i < strs.len) {
		let j: i32 = 0;
		for (j < strs[i].len) {
			buf[off + j] = strs[i][j];
			j += 1;
		};
		off += strs[i].len;
		if (i + 1 < strs.len) {
			j = 0;
			for (j < delim.len) {
				buf[off + j] = delim[j];
				j += 1;
			};
			off += delim.len;
		};
		i += 1;
	};
	r.ptr = buf;
	r.len = total;
	return r;
};

// sub — borrowed `s[start..end]`. ref/hare/strings/sub.ha:30 is
// rune-wise; this ww form is byte-wise (no rune iterator yet, planned
// for commit 2). Clamps out-of-range silently where Hare aborts —
// retained for the existing getopt caller; will graduate when the
// rune-wise form lands.
export fn sub(s: str, start: i32, end: i32) str = {
	let lo: i32 = start;
	let hi: i32 = end;
	if (lo < 0)     { lo = 0; };
	if (hi > s.len) { hi = s.len; };
	if (hi < lo)    { hi = lo; };
	let r: str;
	r.ptr = s.ptr + (lo: u64);
	r.len = hi - lo;
	return r;
};

// runebytes — encode `r` into caller's `scratch` (must hold 4 bytes)
// and return the borrowed slice trimmed to the encoded length. Hare
// inlines the same shape at ref/hare/strings/index.ha:132.
fn runebytes(scratch: []u8, r: rune) []u8 = {
	let n: i32 = utf8.encoderune(scratch, r);
	let s: []u8;
	s.ptr = scratch.ptr;
	s.len = n;
	s.cap = n;
	return s;
};

// hasprefix — true iff `in` begins with `prefix`.
// ref/hare/strings/suffix.ha:8.
export fn hasprefix(in: str, prefix: (str | rune)) bool = {
	let scratch: [4]u8;
	let p: []u8 = match (prefix) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.hasprefix(toutf8(in), p);
};

// hassuffix — true iff `in` ends with `suff`.
// ref/hare/strings/suffix.ha:26.
export fn hassuffix(in: str, suff: (str | rune)) bool = {
	let scratch: [4]u8;
	let s: []u8 = match (suff) {
	case let v: str  => yield toutf8(v);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.hassuffix(toutf8(in), s);
};

// byteindex — byte-wise offset of `needle` in `haystack`, or void if
// absent. ref/hare/strings/index.ha:127. Rune arm encodes via
// utf8.encoderune (Hare passes the encoded slice straight to
// bytes::index).
export fn byteindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	let scratch: [4]u8;
	let n: []u8 = match (needle) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.index(toutf8(haystack), n);
};

// rbyteindex — byte-wise offset of the last `needle` in `haystack`.
// ref/hare/strings/index.ha:138.
export fn rbyteindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	let scratch: [4]u8;
	let n: []u8 = match (needle) {
	case let s: str  => yield toutf8(s);
	case let r: rune => yield runebytes(scratch[0:4], r);
	};
	return bytes.rindex(toutf8(haystack), n);
};

// index — rune-wise offset of `needle`'s first occurrence in
// `haystack`, or void if absent. ref/hare/strings/index.ha:10. The
// str-arm reuses `byteindex` for the anchor byte offset and then
// walks `iter` forward to convert byte→rune index; the rune-arm
// mirrors Hare's `index_rune` (ref/hare/strings/index.ha:31).
export fn index(haystack: str, needle: (str | rune)) (i32 | void) = {
	match (needle) {
	case let s: str  => {
		match (byteindex(haystack, s)) {
		case void        => return;
		case let bo: i32 => {
			let it: iterator = iter(haystack);
			let i: i32 = 0;
			for (position(&it) < bo) {
				match (next(&it)) {
				case let r: rune => i += 1;
				case utf8.done   => break;
				};
			};
			return i;
		};
		};
	};
	case let r: rune => {
		let it: iterator = iter(haystack);
		let i: i32 = 0;
		for (true) {
			match (next(&it)) {
			case let n: rune => {
				if (n == r) { return i; };
				i += 1;
			};
			case utf8.done   => return;
			};
		};
	};
	};
	return;
};

// rindex — rune-wise offset of `needle`'s last occurrence in
// `haystack`, or void if absent. ref/hare/strings/index.ha:22. The
// str-arm reuses `rbyteindex`; the rune-arm walks forward tracking
// the most recent matching rune index (Hare's `rindex_rune` with
// `riter` returns a byte-offset value for multibyte strings, which
// disagrees with the rune-wise docstring; we keep the docstring's
// contract).
export fn rindex(haystack: str, needle: (str | rune)) (i32 | void) = {
	match (needle) {
	case let s: str  => {
		match (rbyteindex(haystack, s)) {
		case void        => return;
		case let bo: i32 => {
			let it: iterator = iter(haystack);
			let i: i32 = 0;
			for (position(&it) < bo) {
				match (next(&it)) {
				case let r: rune => i += 1;
				case utf8.done   => break;
				};
			};
			return i;
		};
		};
	};
	case let r: rune => {
		let it: iterator = iter(haystack);
		let i: i32 = 0;
		let last: i32 = -1;
		for (true) {
			match (next(&it)) {
			case let n: rune => {
				if (n == r) { last = i; };
				i += 1;
			};
			case utf8.done   => break;
			};
		};
		if (last < 0) { return; };
		return last;
	};
	};
	return;
};

// contains — true iff any of `needles` occurs in `haystack`.
// ref/hare/strings/contains.ha:9.
export fn contains(haystack: str, needles: (str | rune)...) bool = {
	let i: i32 = 0;
	for (i < needles.len) {
		match (needles[i]) {
		case let s: str  => {
			match (byteindex(haystack, s)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		case let r: rune => {
			match (byteindex(haystack, r)) {
			case let bo: i32 => return true;
			case void        => void;
			};
		};
		};
		i += 1;
	};
	return false;
};

// trimprefix — `s` with `prefix` stripped from the front, or `s`
// unchanged if it doesn't start with `prefix`. Borrowed view.
// ref/hare/strings/trim.ha:60.
export fn trimprefix(input: str, prefix: str) str = {
	if (!hasprefix(input, prefix)) { return input; };
	let r: str;
	r.ptr = input.ptr + (prefix.len: u64);
	r.len = input.len - prefix.len;
	return r;
};

// trimsuffix — symmetric. ref/hare/strings/trim.ha:69.
export fn trimsuffix(input: str, suffix: str) str = {
	if (!hassuffix(input, suffix)) { return input; };
	let r: str;
	r.ptr = input.ptr;
	r.len = input.len - suffix.len;
	return r;
};

// ltrim — strip leading runes that occur in `trim`. Borrowed view.
// Empty `trim` returns input unchanged (Hare's no-rune branch strips
// ASCII whitespace via `bytes::ltrim`; needs lib/bytes variadic
// graduation). ref/hare/strings/trim.ha:11.
export fn ltrim(input: str, trim: rune...) str = {
	if (trim.len == 0) { return input; };
	let it: iterator = iter(input);
	for (true) {
		match (next(&it)) {
		case let r: rune => {
			let j: i32 = 0;
			let found: bool = false;
			for (j < trim.len) {
				if (r == trim[j]) { found = true; j = trim.len; }
				else { j += 1; };
			};
			if (!found) {
				match (prev(&it)) {
				case let r2: rune => void;
				case utf8.done    => void;
				};
				break;
			};
		};
		case utf8.done => break;
		};
	};
	return iterstr(&it);
};

// rtrim — strip trailing runes that occur in `trim`. Borrowed view.
// ref/hare/strings/trim.ha:32.
export fn rtrim(input: str, trim: rune...) str = {
	if (trim.len == 0) { return input; };
	let it: iterator = riter(input);
	for (true) {
		match (next(&it)) {
		case let r: rune => {
			let j: i32 = 0;
			let found: bool = false;
			for (j < trim.len) {
				if (r == trim[j]) { found = true; j = trim.len; }
				else { j += 1; };
			};
			if (!found) {
				match (prev(&it)) {
				case let r2: rune => void;
				case utf8.done    => void;
				};
				break;
			};
		};
		case utf8.done => break;
		};
	};
	return iterstr(&it);
};

// trim — strip from both ends. ref/hare/strings/trim.ha:54.
export fn trim(input: str, trim: rune...) str = {
	return ltrim(rtrim(input, trim...), trim...);
};

// iterator — UTF-8 rune cursor over a `str`. Layout flattens Hare's
// anonymous-embedded `utf8::decoder` (ref/hare/strings/iter.ha:6-9) to
// explicit fields. `reverse` selects walk direction: forward iterators
// (`iter`) advance through utf8.next; reverse iterators (`riter`) advance
// through utf8.prev. May be copied to save state.
export type iterator = struct {
	offs:    i32,
	src:     []u8,
	reverse: bool,
};

// iter — initialize a forward iterator at the start of `src`.
// ref/hare/strings/iter.ha:24.
export fn iter(src: str) iterator = {
	let r: iterator;
	r.src = toutf8(src);
	r.offs = 0;
	r.reverse = false;
	return r;
};

// riter — initialize a reverse iterator at the end of `src`. `next`
// on a reverse iterator walks back through the string.
// ref/hare/strings/iter.ha:32.
export fn riter(src: str) iterator = {
	let r: iterator;
	r.src = toutf8(src);
	r.offs = src.len;
	r.reverse = true;
	return r;
};

// move — private dispatch shared by next/prev. `forward` selects
// utf8.next vs utf8.prev. Aborts on more/invalid per Hare's
// ref/hare/strings/iter.ha:51-58 ("Invalid UTF-8 string (this should
// not happen)"). Hare picks the utf8 function via a fn-pointer; ww
// branches on `forward` at each call site instead.
fn move(forward: bool, it: *iterator) (rune | utf8.done) = {
	let d: utf8.decoder;
	d.src = it.src;
	d.offs = it.offs;
	if (forward) {
		match (utf8.next(&d)) {
		case let r: rune          => { it.offs = d.offs; return r; };
		case let dn: utf8.done    => return dn;
		case let m:  utf8.more    => abort("strings.move: invalid UTF-8");
		case let e:  utf8.invalid => abort("strings.move: invalid UTF-8");
		};
	} else {
		match (utf8.prev(&d)) {
		case let r: rune          => { it.offs = d.offs; return r; };
		case let dn: utf8.done    => return dn;
		case let m:  utf8.more    => abort("strings.move: invalid UTF-8");
		case let e:  utf8.invalid => abort("strings.move: invalid UTF-8");
		};
	};
};

// next — advance the iterator one rune. Forward iterators step
// through utf8.next; reverse iterators (riter) step backward through
// utf8.prev. Returns utf8.done at end-of-walk. ref/hare/strings/iter.ha:45.
export fn next(it: *iterator) (rune | utf8.done) = {
	return move(!it.reverse, it);
};

// prev — step back one rune. Dual to next: on a forward iterator
// this walks utf8.prev; on a reverse iterator (riter) it walks
// utf8.next. ref/hare/strings/iter.ha:49.
export fn prev(it: *iterator) (rune | utf8.done) = {
	return move(it.reverse, it);
};

// iterstr — borrowed view of the bytes remaining in the iterator's
// walk direction. Forward iter: bytes from offs to end; reverse iter:
// bytes from start to offs. ref/hare/strings/iter.ha:63.
export fn iterstr(it: *iterator) str = {
	let r: []u8;
	if (it.reverse) {
		r = it.src[0:it.offs];
	} else {
		r = it.src[it.offs:it.src.len];
	};
	return fromutf8_unsafe(r);
};

// slice — borrowed substring between two iterator positions.
// ref/hare/strings/iter.ha:75. Hare passes `*iterator` directly where
// `*utf8::decoder` is expected via anonymous-embed coercion; ww has
// no anonymous embed, so we reconstruct a local utf8.decoder for each
// endpoint and forward — same pattern as `move` above.
export fn slice(begin: *iterator, end: *iterator) str = {
	let b: utf8.decoder;
	b.src = begin.src;
	b.offs = begin.offs;
	let e: utf8.decoder;
	e.src = end.src;
	e.offs = end.offs;
	return fromutf8_unsafe(utf8.slice(&b, &e));
};

// position — byte-wise offset of the iterator in its source.
// ref/hare/strings/iter.ha:82.
export fn position(it: *iterator) i32 = {
	return it.offs;
};

// tokenizer — re-export of bytes.tokenizer. ref/hare/strings/tokenize.ha:7.
// First cross-module type alias in tree; needs #22's transitive
// alias-chain unwrap (cstage type_chase_named + wwstage
// structlookupchain) to walk struct fields through the chain.
export type tokenizer = bytes.tokenizer;

// tokenize — yield substrings of `s` split on any byte in `delim`.
// Leading / trailing / adjacent delims yield empty tokens. `s` and
// `delim` are borrowed; caller keeps them live for the tokenizer's
// lifetime. ref/hare/strings/tokenize.ha:32. ASCII-only delim
// asserted per Hare lines 35-37: a multibyte rune in delim would
// split on a single continuation byte and yield invalid UTF-8.
export fn tokenize(s: str, delim: str) tokenizer = {
	let d: []u8 = toutf8(delim);
	let i: i32 = 0;
	for (i < d.len) {
		os.assert((d[i] & 0x80u8) == 0u8,
			"strings.tokenize cannot tokenize on non-ASCII delimiters");
		i += 1;
	};
	return bytes.tokenize(toutf8(s), d...);
};

// rtokenize — reverse-direction counterpart to [[tokenize]]. First
// next_token yields the last token, last yields the first.
// ref/hare/strings/tokenize.ha:44.
export fn rtokenize(s: str, delim: str) tokenizer = {
	let d: []u8 = toutf8(delim);
	let i: i32 = 0;
	for (i < d.len) {
		os.assert((d[i] & 0x80u8) == 0u8,
			"strings.rtokenize cannot tokenize on non-ASCII delimiters");
		i += 1;
	};
	return bytes.rtokenize(toutf8(s), d...);
};

// next_token — current token, advancing the cursor.
// ref/hare/strings/tokenize.ha:62.
export fn next_token(s: *tokenizer) (str | bytes.done) = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	match (bytes.next_token(b)) {
	case let v: []u8 => return fromutf8_unsafe(v);
	case bytes.done  => { let d: bytes.done; return d; };
	};
};

// peek_token — current token without advancing.
// ref/hare/strings/tokenize.ha:71.
export fn peek_token(s: *tokenizer) (str | bytes.done) = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	match (bytes.peek_token(b)) {
	case let v: []u8 => return fromutf8_unsafe(v);
	case bytes.done  => { let d: bytes.done; return d; };
	};
};

// remaining_tokens — unconsumed portion of the input ahead of the
// cursor. ref/hare/strings/tokenize.ha:79.
export fn remaining_tokens(s: *tokenizer) str = {
	let b: *bytes.tokenizer = s: *bytes.tokenizer;
	return fromutf8_unsafe(bytes.remaining_tokens(b));
};

// rt_ensure is the runtime slice-growth helper invoked by the
// `append(s, v)` builtin. Direct bind for the same reason as
// lib/shlex.shlex (appendstr, 16B): the builtin's expansion stores
// only 8B of the new element, losing the `.len` half of a `str`.
@symbol("rt_ensure") fn rtensure(s: *void, membsz: u64) void;

// appendstr — grow `*slice` by one and store `item` (16B). Mirror of
// lib/shlex.shlex appendstr. Collapses when the append builtin learns
// to store the full element width.
fn appendstr(slice: *[]str, item: str) void = {
	let newlen: i32 = slice.len + 1;
	slice.len = newlen;
	rtensure(slice: *void, 16u64);
	let dst: *str = &slice.ptr[newlen - 1];
	dst.ptr = item.ptr;
	dst.len = item.len;
};

// splitn — split `in` on any byte in `delim`, returning up to `n`
// tokens via forward iteration. The trailing slot (when more than
// `n - 1` tokens exist) holds the unconsumed remainder. Strings
// within the result are borrowed from `in`.
//
// The caller frees the returned slice via
// `os.free(r.ptr: *void, (r.cap: u64) * 16u64)`.
//
// Hare's `([]str | nomem)` collapses to `[]str` here: ww os.alloc
// has no recoverable failure path. Same precedent as
// shlex.split / bytes.splitn.
//
// ref/hare/strings/tokenize.ha:172.
export fn splitn(in: str, delim: str, n: i32) []str = {
	let toks: []str;
	toks.ptr = nil: *str;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = tokenize(in, delim);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: str => { appendstr(&toks, s); };
		case bytes.done => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case bytes.done  => void;
	case let pk: str => {
		let r: str = remaining_tokens(&tok);
		appendstr(&toks, r);
	};
	};
	return toks;
};

// rsplitn — reverse-direction counterpart to [[splitn]]: tokens are
// collected from the end of `in`. The trailing slot holds the
// unconsumed prefix (everything before the n-th-from-last delim hit).
//
// When the input has fewer than n tokens, the `done` short-circuit
// returns toks UN-reversed (in last-token-first order). Mirrors Hare
// at ref/hare/strings/tokenize.ha:219-224 where the in-place reverse
// step is gated behind the n-1 loop running to completion.
//
// ref/hare/strings/tokenize.ha:200.
export fn rsplitn(in: str, delim: str, n: i32) []str = {
	let toks: []str;
	toks.ptr = nil: *str;
	toks.len = 0;
	toks.cap = 0;
	let tok: tokenizer = rtokenize(in, delim);
	let i: i32 = 0;
	for (i < n - 1) {
		match (next_token(&tok)) {
		case let s: str => { appendstr(&toks, s); };
		case bytes.done => { return toks; };
		};
		i += 1;
	};
	match (peek_token(&tok)) {
	case bytes.done  => void;
	case let pk: str => {
		let r: str = remaining_tokens(&tok);
		appendstr(&toks, r);
	};
	};

	// In-place reverse so callers see argv-order, matching Hare
	// (ref/hare/strings/tokenize.ha:220). Element copy is field-wise
	// through `*str` because `toks[i] = toks[j]` (full 16B str store)
	// lands in the multi-word-store gap noted at cmd/w6c/cgen.c:6515.
	let a: i32 = 0;
	let b: i32 = toks.len - 1;
	for (a < b) {
		let pa: *str = &toks.ptr[a];
		let pb: *str = &toks.ptr[b];
		let tp: *u8 = pa.ptr;
		let tl: i32 = pa.len;
		pa.ptr = pb.ptr;
		pa.len = pb.len;
		pb.ptr = tp;
		pb.len = tl;
		a += 1;
		b -= 1;
	};
	return toks;
};

// split — full split of `in` on `delim` (no token cap). Mirrors
// `splitn(in, delim, types::SIZE_MAX)`. ww uses `types.I32_MAX`
// because the index type is i32 (lib/CLAUDE.md).
//
// ref/hare/strings/tokenize.ha:242.
export fn split(in: str, delim: str) []str = {
	return splitn(in, delim, types.I32_MAX);
};

// lpad — left-pad `s` with `p` rune until the result reaches `maxlen`
// bytes. Length comparison is BYTES, mirroring Hare's `len(s) >= maxlen`
// at ref/hare/strings/pad.ha:9. A multibyte `p` whose encoded width
// doesn't divide `maxlen - s.len` evenly leaves a trailing pad byte
// pair sliced mid-codepoint at byte `maxlen-1`, exactly as Hare's
// `res[..maxlen]` does (ref/hare/strings/pad.ha:20). When
// `(maxlen - s.len) * pad.len >= maxlen` (multibyte pad overflows the
// budget), `s` is entirely sliced off — same as Hare. Caller releases
// with `os.free(r.ptr, r.len: u64)`. Hare's `nomem` return is dropped:
// `os.alloc` aborts on OOM. Buf size == r.len keeps the free-contract
// shape of [[dup]] / [[concat]] / [[join]]; Hare's `alloc([], maxlen)!`
// over-allocs via append then slices, but Hare's slice-free recovers
// the true capacity from the heap allocator (rt/ensure.ha:24), which
// ww's munmap-based `os.free` cannot do.
export fn lpad(s: str, p: rune, maxlen: i32) str = {
	if (s.len >= maxlen) { return dup(s); };
	let scratch: [4]u8;
	let pad: []u8 = runebytes(scratch[0:4], p);
	let buf: *u8 = os.alloc(maxlen: u64): *u8;
	let padwrite: i32 = (maxlen - s.len) * pad.len;
	if (padwrite > maxlen) { padwrite = maxlen; };
	let off: i32 = 0;
	for (off < padwrite) {
		buf[off] = pad.ptr[off % pad.len];
		off += 1;
	};
	let k: i32 = 0;
	let srem: i32 = maxlen - off;
	if (srem > s.len) { srem = s.len; };
	for (k < srem) {
		buf[off + k] = s[k];
		k += 1;
	};
	let r: str;
	r.ptr = buf;
	r.len = maxlen;
	return r;
};

// rpad — right-pad `s` with `p` rune until the result reaches `maxlen`
// bytes. Symmetric with [[lpad]]. ref/hare/strings/pad.ha:39.
export fn rpad(s: str, p: rune, maxlen: i32) str = {
	if (s.len >= maxlen) { return dup(s); };
	let scratch: [4]u8;
	let pad: []u8 = runebytes(scratch[0:4], p);
	let buf: *u8 = os.alloc(maxlen: u64): *u8;
	let k: i32 = 0;
	for (k < s.len) {
		buf[k] = s[k];
		k += 1;
	};
	let padwrite: i32 = maxlen - s.len;
	let i: i32 = 0;
	for (i < padwrite) {
		buf[s.len + i] = pad.ptr[i % pad.len];
		i += 1;
	};
	let r: str;
	r.ptr = buf;
	r.len = maxlen;
	return r;
};

// strconv — number↔string conversions.
//
// Mirrors Hare's strconv:: surface. The *tos functions return a
// `const str` view into a module-level buffer that is overwritten on
// the next call to the same function; callers must copy the bytes if
// they need to outlive the next invocation. See [[strings.dup]] to
// duplicate. Matches Hare's strconv::*tos semantics.

package strconv;

import os;
import strings;

// invalid — input wasn't a valid number in the requested format.
// Payload is the byte index of the first offending position.
// Mirrors Hare's strconv::invalid = !size.
export type invalid = !i32;

// overflow — input was valid but doesn't fit the target type.
// Mirrors Hare's strconv::overflow = !void.
export type overflow = !void;

// error — any error from a strconv call. Mirrors Hare's strconv::error.
export type error = !(invalid | overflow);

// base — numeric base for parsing/formatting. Mirrors Hare's
// `strconv::base` (Hare uses `enum uint`; we pick `enum i32` since
// the underlying parse/format loops index with i32).
//
// HEX is an alias for HEX_UPPER; HEX_LOWER is a pseudo-base that
// produces lowercase a-f digits.
export type base = enum i32 {
	DEFAULT   = 0,
	BIN       = 2,
	OCT       = 8,
	DEC       = 10,
	HEX_UPPER = 16,
	HEX       = 16,
	HEX_LOWER = 17,
};

fn basenum(b: base) i64 = {
	if (b == base.BIN)       { return 2;  };
	if (b == base.OCT)       { return 8;  };
	if (b == base.HEX)       { return 16; };
	if (b == base.HEX_UPPER) { return 16; };
	if (b == base.HEX_LOWER) { return 16; };
	return 10;	// DEC and DEFAULT
};

fn basedigit(d: i64, b: base) u8 = {
	if (d < 10) { return (d + 48): u8; };
	let off: i64 = d - 10;
	if (b == base.HEX_LOWER) { return (off + 97): u8; };
	return (off + 65): u8;
};

// u64tos — convert v to a base-b numeric string. Returns a view into
// `u64tos_buf` which is overwritten on the next call. Matches Hare's
// strconv::u64tos.
let u64tos_buf: [65]u8;

export fn u64tos(v: u64, b: base) str = {
	let nb: u64 = basenum(b): u64;
	let tmp: [65]u8;
	let i: i32 = 0;
	let n: u64 = v;
	if (n == 0u64) { tmp[0] = 48u8; i = 1; };
	for (n > 0u64) {
		let d: i64 = (n % nb): i64;
		tmp[i] = basedigit(d, b);
		n = n / nb;
		i += 1;
	};
	let out: i32 = 0;
	for (i > 0) {
		i -= 1;
		u64tos_buf[out] = tmp[i];
		out += 1;
	};
	let r: str;
	r.ptr = &u64tos_buf[0];
	r.len = out;
	return r;
};

// i64tos — convert v to a base-b numeric string. Returns a view into
// `i64tos_buf` which is overwritten on the next call. Independent
// buffer from u64tos so i64tos's own call to u64tos doesn't clobber
// the in-flight result. Matches Hare's strconv::i64tos.
let i64tos_buf: [66]u8;

export fn i64tos(v: i64, b: base) str = {
	let neg: bool = false;
	let n: i64 = v;
	if (n < 0) { neg = true; n = -n; };
	let nb: i64 = basenum(b);
	let tmp: [65]u8;
	let i: i32 = 0;
	if (n == 0) { tmp[0] = 48u8; i = 1; };
	for (n > 0) {
		let d: i64 = n % nb;
		tmp[i] = basedigit(d, b);
		n = n / nb;
		i += 1;
	};
	let out: i32 = 0;
	if (neg) { i64tos_buf[out] = 45u8; out += 1; };	// '-'
	for (i > 0) {
		i -= 1;
		i64tos_buf[out] = tmp[i];
		out += 1;
	};
	let r: str;
	r.ptr = &i64tos_buf[0];
	r.len = out;
	return r;
};

export fn i32tos(v: i32, b: base) str = { return i64tos(v: i64, b); };
export fn i16tos(v: i16, b: base) str = { return i64tos(v: i64, b); };
export fn i8tos(v: i8,   b: base) str = { return i64tos(v: i64, b); };

export fn u32tos(v: u32, b: base) str = { return u64tos(v: u64, b); };
export fn u16tos(v: u16, b: base) str = { return u64tos(v: u64, b); };
export fn u8tos(v: u8,   b: base) str = { return u64tos(v: u64, b); };

// digval — value of digit byte `c` under base `b`, or -1 if not a
// valid digit. Letters are accepted case-insensitively under HEX /
// HEX_UPPER; only lowercase under HEX_LOWER.
fn digval(c: u8, b: base) i32 = {
	if (c >= 48u8) { if (c <= 57u8) { return (c - 48u8): i32; }; };
	if (b == base.HEX_LOWER) {
		if (c >= 97u8) { if (c <= 102u8) { return ((c - 97u8) + 10u8): i32; }; };
		return -1;
	};
	if (c >= 65u8) { if (c <= 70u8)  { return ((c - 65u8) + 10u8): i32; }; };
	if (c >= 97u8) { if (c <= 102u8) { return ((c - 97u8) + 10u8): i32; }; };
	return -1;
};

// stoi64 — parse signed base-b number. Mirrors Hare's strconv::stoi64.
// No locale, no whitespace, no underscores: optional leading '-' then
// digits. Returns invalid with the offending index or overflow on
// out-of-range.
export fn stoi64(s: str, b: base) (i64 | invalid | overflow) = {
	if (s.len == 0) { return 0: invalid; };
	let i: i32 = 0;
	let neg: bool = false;
	if (s[0] == 45u8) { neg = true; i = 1; };
	if (i >= s.len) { return i: invalid; };
	let nb: i32 = basenum(b): i32;
	let v: i64 = 0;
	for (i < s.len) {
		let c: u8 = s[i];
		let d: i32 = digval(c, b);
		if (d < 0)   { return i: invalid; };
		if (d >= nb) { return i: invalid; };
		v = v * (nb: i64) + (d: i64);
		i += 1;
	};
	if (neg) { v = -v; };
	return v;
};

// stou64 — parse unsigned base-b number. Mirrors Hare's strconv::stou64.
export fn stou64(s: str, b: base) (u64 | invalid | overflow) = {
	if (s.len == 0) { return 0: invalid; };
	let nb: u64 = basenum(b): u64;
	let v: u64 = 0u64;
	let i: i32 = 0;
	for (i < s.len) {
		let c: u8 = s[i];
		let d: i32 = digval(c, b);
		if (d < 0)         { return i: invalid; };
		if ((d: u64) >= nb) { return i: invalid; };
		v = v * nb + (d: u64);
		i += 1;
	};
	return v;
};

export fn stoi32(s: str, b: base) (i32 | invalid | overflow) = {
	let r = stoi64(s, b);
	match (r) {
	case let v: i64 => {
		if (v > 2147483647i64)  { return overflow{}; };
		if (v < -2147483648i64) { return overflow{}; };
		return v: i32;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;	// unreachable; appeases the path-cov checker
};

export fn stoi16(s: str, b: base) (i16 | invalid | overflow) = {
	let r = stoi64(s, b);
	match (r) {
	case let v: i64 => {
		if (v > 32767i64)  { return overflow{}; };
		if (v < -32768i64) { return overflow{}; };
		return v: i16;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;
};

export fn stoi8(s: str, b: base) (i8 | invalid | overflow) = {
	let r = stoi64(s, b);
	match (r) {
	case let v: i64 => {
		if (v > 127i64)  { return overflow{}; };
		if (v < -128i64) { return overflow{}; };
		return v: i8;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;
};

export fn stou32(s: str, b: base) (u32 | invalid | overflow) = {
	let r = stou64(s, b);
	match (r) {
	case let v: u64 => {
		if (v > 4294967295u64) { return overflow{}; };
		return v: u32;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;
};

export fn stou16(s: str, b: base) (u16 | invalid | overflow) = {
	let r = stou64(s, b);
	match (r) {
	case let v: u64 => {
		if (v > 65535u64) { return overflow{}; };
		return v: u16;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;
};

export fn stou8(s: str, b: base) (u8 | invalid | overflow) = {
	let r = stou64(s, b);
	match (r) {
	case let v: u64 => {
		if (v > 255u64) { return overflow{}; };
		return v: u8;
	};
	case let e: invalid  => return e;
	case let e: overflow => return e;
	};
	return 0: invalid;
};

// f64tos — convert v to a decimal string. Returns owned str; release
// via os.free. Mirrors Hare's strconv::f64tos (current ww impl is
// fixed-point only, max 6 fractional digits, no NaN/Inf support —
// see graduate-to-Ryū note below).
//
// Surface:
//
//   - finite values only. NaN/±Inf detection needs an f64→u64 bit
//     reinterpret cast that the cgen doesn't expose yet.
//   - fixed-point only, up to 6 fractional digits. Trailing zeros
//     after the decimal point are trimmed. Trailing '.' is dropped.
//   - magnitudes ≥ 9e18 (overflows i64 in the integer-part cast)
//     fall back to the literal token "huge". Hare would print these
//     in scientific notation via Ryū; we will graduate when the
//     compiler grows the bit-reinterpret cast.
//
// Round-trip is therefore lossy past 6 fractional digits.
//
// No float literals in the body — 990's wwdump diff requires this
// file's TK_FLOAT count to match between C and ww front-ends, and
// the ww-side wwdump currently skips TK_FLOAT.fval while the C side
// %g-formats it. Same trick lib/ww/lex/lex.ww's parsef64 uses:
// build f64 constants via int-to-f64 casts.
let f64tos_buf: [64]u8;

export fn f64tos(v: f64) str = {
	let out: i32 = 0;
	let f: f64 = v;
	let zero: f64 = 0: f64;
	if (f < zero) {
		f64tos_buf[out] = 45u8;	// '-'
		out += 1;
		f = -f;
	};
	// 9e18 is comfortably under I64_MAX (9.22e18). Past this the
	// `f: i64` cast wraps and the integer part comes back as garbage.
	let cap: f64 = 9000000000000000000i64: f64;
	if (f >= cap) {
		let s: str = "huge";
		let k: i32 = 0;
		for (k < s.len) { f64tos_buf[out] = s[k]; out += 1; k += 1; };
		let r: str;
		r.ptr = &f64tos_buf[0];
		r.len = out;
		return r;
	};
	let ip: i64 = f: i64;
	// Fractional part scaled to 6 decimal digits, with round-to-
	// nearest via +0.5. (f64 compound assigns mis-lower in cgen —
	// use the explicit form, as the rest of lib does.)
	let frac: f64 = f - (ip: f64);
	let scale: f64 = 1000000: f64;
	frac = frac * scale;
	let half: f64 = (1: f64) / (2: f64);
	let fp: i64 = (frac + half): i64;
	// Carry: e.g. 0.9999996 rounds fp up to 1000000 and the integer
	// part needs to advance.
	if (fp >= 1000000) {
		ip += 1;
		fp = 0;
	};
	let intstr: str = i64tos(ip, base.DEC);
	let k: i32 = 0;
	for (k < intstr.len) { f64tos_buf[out] = intstr.ptr[k]; out += 1; k += 1; };
	if (fp != 0) {
		f64tos_buf[out] = 46u8;	// '.'
		out += 1;
		let fracstr: str = u64tos(fp: u64, base.DEC);
		// Pad fractional to 6 digits with leading zeros (e.g. 0.05 →
		// fp=50000, fracstr="50000", pad one '0' before).
		let z: i32 = 6 - fracstr.len;
		for (z > 0) { f64tos_buf[out] = 48u8; out += 1; z -= 1; };
		k = 0;
		for (k < fracstr.len) { f64tos_buf[out] = fracstr.ptr[k]; out += 1; k += 1; };
		// Trim trailing zeros in the fractional part.
		for (out > 0) {
			if (f64tos_buf[out - 1] != 48u8) { break; };
			out -= 1;
		};
	};
	let r: str;
	r.ptr = &f64tos_buf[0];
	r.len = out;
	return r;
};

// strerror — convert an strconv error to a user-readable string.
// Returns owned str; release via os.free. Mirrors Hare's
// strconv::strerror.
export fn strerror(e: error) str = {
	match (e) {
	case let v: invalid  => return strings.dup("input is not a valid number");
	case let v: overflow => return strings.dup("input number doesn't fit target type");
	};
	return strings.dup("");
};

// ascii — rune-class predicates and case folding for the ASCII range.
// Matches Hare's ascii::isdigit family (rune-taking signature). Runes
// outside 0..127 always answer `false`. The lexer hot path uses these
// inline; they are expected to inline to a couple of compares.

package ascii;

export fn isdigit(c: rune) bool = {
	if (c < 48) { return false; };
	if (c > 57) { return false; };
	return true;
};

export fn isupper(c: rune) bool = {
	if (c < 65) { return false; };
	if (c > 90) { return false; };
	return true;
};

export fn islower(c: rune) bool = {
	if (c < 97) { return false; };
	if (c > 122) { return false; };
	return true;
};

export fn isalpha(c: rune) bool = {
	if (isupper(c)) { return true; };
	return islower(c);
};

export fn isalnum(c: rune) bool = {
	if (isalpha(c)) { return true; };
	return isdigit(c);
};

// isspace — the C/Hare set: space, tab, NL, VT, FF, CR.
export fn isspace(c: rune) bool = {
	if (c == 32) { return true; };	// ' '
	if (c == 9)  { return true; };	// '\t'
	if (c == 10) { return true; };	// '\n'
	if (c == 11) { return true; };	// '\v'
	if (c == 12) { return true; };	// '\f'
	if (c == 13) { return true; };	// '\r'
	return false;
};

export fn isxdigit(c: rune) bool = {
	if (isdigit(c)) { return true; };
	if (c >= 65) {
		if (c <= 70) { return true; };	// 'A'..'F'
	};
	if (c >= 97) {
		if (c <= 102) { return true; };	// 'a'..'f'
	};
	return false;
};

// valid — `c` is in the 0..127 ASCII range.
export fn valid(c: rune) bool = {
	if (c < 0)   { return false; };
	if (c > 127) { return false; };
	return true;
};

// validstr — every byte in `s` is ASCII (0..127).
export fn validstr(s: str) bool = {
	let i: i32 = 0;
	for (i < s.len) {
		// High-bit test rather than `> 127u8`; both cgens lower
		// the bitwise form identically. The `> u8` form picks
		// JA vs JG depending on signed/unsigned dispatch.
		if ((s[i] & 128u8) != 0u8) { return false; };
		i += 1;
	};
	return true;
};

// iscntrl — control chars: 0..31 and 127.
export fn iscntrl(c: rune) bool = {
	if (c >= 0) { if (c <= 31) { return true; }; };
	if (c == 127) { return true; };
	return false;
};

// isblank — space and tab.
export fn isblank(c: rune) bool = {
	if (c == 32) { return true; };	// ' '
	if (c == 9)  { return true; };	// '\t'
	return false;
};

// isprint — printable: space through '~'.
export fn isprint(c: rune) bool = {
	if (c < 32)  { return false; };
	if (c > 126) { return false; };
	return true;
};

// isgraph — printable, non-space.
export fn isgraph(c: rune) bool = {
	if (c < 33)  { return false; };
	if (c > 126) { return false; };
	return true;
};

// ispunct — printable, non-alnum, non-space.
export fn ispunct(c: rune) bool = {
	if (!isgraph(c)) { return false; };
	if (isalnum(c))  { return false; };
	return true;
};

// tolower / toupper — fold ASCII case. Non-letters pass through.
export fn tolower(c: rune) rune = {
	if (isupper(c)) { return c + 32; };
	return c;
};

export fn toupper(c: rune) rune = {
	if (islower(c)) { return c - 32; };
	return c;
};

// strcasecmp — three-way ASCII case-insensitive compare.
export fn strcasecmp(a: str, b: str) i32 = {
	let n: i32 = a.len;
	if (b.len < n) { n = b.len; };
	let i: i32 = 0;
	for (i < n) {
		let ca: rune = tolower(a[i]: rune);
		let cb: rune = tolower(b[i]: rune);
		if (ca != cb) { return (ca - cb): i32; };
		i += 1;
	};
	return a.len - b.len;
};

// selfhost/test/smoke.ww — end-to-end smoke for the selfhost path.
//
// Exercises the patterns the real ww-side compiler port will use:
//   - bump arena allocator (mem.ww shape)
//   - error idiom (T | str)
//   - struct of fn pointers + ctx pointer (the io.stream-style
//     polymorphism we use instead of interfaces)
//   - byte-level scanning that mirrors the hot path inside lex.ww
//   - strconv round-trip via the real stdlib
//
// `main` returns 42 when every check passes, 1..N on failure
// indicating which probe broke. The 990_selfhost test asserts 42.
//
// Note: only stack-local mutable state. Top-level `let` mutation
// requires a writable .data segment in w6l, which is a separate
// task; until then we exercise polymorphism via ctx pointers, which
// is what the real port wants anyway.

package test;

import os;
import strconv;
import ascii;

// --- bump arena ---------------------------------------------------------

type arena = struct {
	buf: *u8,
	off: u64,
	cap: u64,
};

// In-place init. Returning a 24-byte struct by value isn't yet
// supported in w6c (SysV requires a hidden return-slot pointer for
// structs >16 bytes), so we initialize through a pointer like the
// real compiler does today.
fn arena_init(a: *arena, buf: *u8, cap: u64) void = {
	a.buf = buf;
	a.off = 0u64;
	a.cap = cap;
};

fn arena_alloc(a: *arena, n: u64) *u8 = {
	if (n > a.cap - a.off) { return nil; };
	let p: *u8 = a.buf + a.off;
	a.off += n;
	return p;
};

// --- (i32 | str) error idiom -------------------------------------------

fn checked_div(num: i32, den: i32) (i32 | str) = {
	if (den == 0) { return "div by zero"; };
	return num / den;
};

// --- struct-of-fn-pointer polymorphism ---------------------------------
//
// A trivial "writer" abstraction: a function pointer plus a context.
// This mirrors how io.stream / Plan 9 Bio work. The ctx pointer lets
// the implementation own its own state without a global.

type counter = struct {
	n: i32,
};

type writer = struct {
	ctx:  *void,
	emit: fn(ctx: *void, b: u8) void,
};

fn count_emit(ctx: *void, b: u8) void = {
	let c: *counter = ctx: *counter;
	c.n += 1;
};

// --- byte scanner like lex.ww's hot path -------------------------------

fn count_digits(s: str) i32 = {
	let i: i32 = 0;
	let n: i32 = 0;
	for (i < s.len) {
		let c: u8 = s[i];
		if (c >= 48u8) {
			if (c <= 57u8) { n += 1; };
		};
		i += 1;
	};
	return n;
};

// --- entry --------------------------------------------------------------

export fn main() i32 = {
	// Probe 1 — arena hands out distinct pointers, refuses oversize.
	let buf: [256]u8;
	let a: arena;
	arena_init(&a, buf.ptr, 256u64);
	let p1: *u8 = arena_alloc(&a, 32u64);
	let p2: *u8 = arena_alloc(&a, 32u64);
	if (p1 == nil) { return 1; };
	if (p2 == nil) { return 2; };
	if (p1 == p2) { return 3; };
	let p3: *u8 = arena_alloc(&a, 1024u64);
	if (p3 != nil) { return 4; };

	// Probe 2 — error union both ways.
	let r_ok: (i32 | str) = checked_div(84, 2);
	let r_bad: (i32 | str) = checked_div(1, 0);
	let acc: i32 = 0;
	match (r_ok) {
		case let v: i32 => acc = v;
		case let e: str => return 5;
	};
	if (acc != 42) { return 6; };
	match (r_bad) {
		case let v: i32 => return 7;
		case let e: str => acc = e.len: i32;
	};
	if (acc != 11) { return 8; };	// len("div by zero") == 11

	// Probe 3 — struct-of-fn-pointer dispatch via ctx pointer.
	let c: counter = counter { n = 0 };
	let w: writer = writer { ctx = (&c): *void, emit = count_emit };
	w.emit(w.ctx, 65u8);
	w.emit(w.ctx, 66u8);
	w.emit(w.ctx, 67u8);
	if (c.n != 3) { return 9; };

	// Probe 4 — byte scan over a literal.
	let dn: i32 = count_digits("ww123abc");
	if (dn != 3) { return 10; };

	// Probe 5 — strconv round-trip via the real stdlib.
	let s: str = strconv.i64tos(4242i64, strconv.base.DEC);
	if (s.len != 4) { return 11; };
	if (s.ptr[0] != 52u8) { return 12; };	// '4'
	if (s.ptr[3] != 50u8) { return 13; };	// '2'

	// Probe 6 — ascii classifications (rune-taking, Hare-shaped).
	if (!ascii.isdigit(53)) { return 14; };		// '5'
	if (ascii.isdigit(65))  { return 15; };		// 'A' is not a digit
	if (!ascii.isalpha(122)) { return 16; };	// 'z'
	if (!ascii.isxdigit(70)) { return 17; };	// 'F'
	if (ascii.isxdigit(71))  { return 18; };	// 'G' is not hex
	if (ascii.tolower(65) != 97) { return 19; };	// 'A' -> 'a'
	if (ascii.toupper(122) != 90) { return 20; };	// 'z' -> 'Z'

	// Probe 7 — file open/read via the new os APIs. /proc/self/cmdline
	// always exists on Linux, no write side, and is non-empty.
	let path: str = "/proc/self/cmdline";
	// Use raw os.open here (returns i32 with -errno) for the same
	// reason as os.read below: probe 6 in 990_selfhost compiles
	// smoke.ww standalone (no `use` expansion), so cross-module type
	// references like `os.oserror` and `os.flag` don't resolve at
	// that step. RDONLY is 0; passing the literal keeps the call
	// site standalone-compilable to byte-identical asm on both
	// compilers.
	let fd: i32 = os.open(path, 0, 0i32);
	if (fd < 0) { return 21; };
	let rbuf: [128]u8;
	// Use raw os.read here (single syscall, plain i64) instead of
	// os.readall: the 990 cgen-match probe compiles smoke.ww
	// standalone without `use os;` expansion, so cross-module type
	// references like `os.oserror` can't be resolved.
	let n: i64 = os.read(fd, rbuf.ptr, 128u64);
	os.close(fd);
	if (n <= 0i64) { return 22; };

	return 42;
};