Make various preparations for upcoming features

A lot of this is random WIP from a while back, at least a month ago, and
is being committed now to get it out of the way so that other patches
can be brought in and integrated against it without causing headaches.

Also rolled into this commit is a way to distinguish plugin_unload from
exiting the game. This is required for another soon-to-be-integrated
feature to avoid crashing on exit, and could in theory also be used to
speed up unloading on exit in future. While we're at it, this also
avoids the need to linearly scan through the plugin list to do the
old branch unloading fix, because we can.

Rough summary of the other smaller stuff I can remember doing:
- Rework bitbuf a bit
- Add some cryptographic nonsense in ac.c (not final at all)
- Introduce the first couple of "chunklets" libraries as a sort-of
  subproject of this one
- Tidy up random small bits and bobs
- Add source for a small keypair generation tool
- Rework democustom to be very marginally more useful

28 files changed, 1767 insertions, 223 deletions

diff --git a/compile b/compile
index bf01493..a94d8b1 100755
--- a/compile
+++ b/compile
@@ -50,11 +50,13 @@ ld() {
 
 src="\
 	ac.c
-	bind.c
-	crypto.c
 	alias.c
 	autojump.c
+	bind.c
+	chunklets/fastspin.c
+	chunklets/msg.c
 	con_.c
+	crypto.c
 	democustom.c
 	demorec.c
 	engineapi.c
diff --git a/compile.bat b/compile.bat
index 75ac2fe..9d8a2a3 100644
--- a/compile.bat
+++ b/compile.bat
@@ -58,11 +58,13 @@ setlocal EnableDelayedExpansion
 for /f "tokens=2" %%f in ('findstr /B /C:":+ " "%~nx0"') do set src=!src! src/%%f

 setlocal DisableDelayedExpansion

 :+ ac.c

-:+ bind.c

-:+ crypto.c

 :+ alias.c

 :+ autojump.c

+:+ bind.c

 :+ con_.c

+:+ chunklets/fastspin.c

+:+ chunklets/msg.c

+:+ crypto.c

 :+ democustom.c

 :+ demorec.c

 :+ engineapi.c

@@ -110,7 +112,7 @@ if "%dbg%"=="1" (
 )

 %CC% -shared -flto %ldflags% -Wl,/IMPLIB:.build/sst.lib,/Brepro,/nodefaultlib ^

 -L.build %clibs% -lkernel32 -luser32 -ladvapi32 -lshlwapi -ld3d9 -ldsound ^

--ltier0 -lvstdlib -o sst.dll%objs% .build/dll.res || exit /b

+-ltier0 -lvstdlib -lntdll -o sst.dll%objs% .build/dll.res || exit /b

 :: get rid of another useless file (can we just not create this???)

 del .build\sst.lib

 

diff --git a/src/ac.c b/src/ac.c
index 605a2be..263e114 100644
--- a/src/ac.c
+++ b/src/ac.c
@@ -20,14 +20,21 @@
 #include <immintrin.h>
 #endif
 
+#include "alias.h"
 #include "bind.h"
+#include "chunklets/fastspin.h"
+#include "chunklets/msg.h"
 #include "con_.h"
+#include "crypto.h"
+#include "democustom.h"
+#include "demorec.h"
 #include "hook.h"
 #include "engineapi.h"
 #include "errmsg.h"
 #include "event.h"
 #include "feature.h"
 #include "gamedata.h"
+#include "gametype.h"
 #include "intdefs.h"
 #include "mem.h"
 #include "os.h"
@@ -37,33 +44,91 @@
 #include "x86.h"
 #include "x86util.h"
 
+#ifdef _WIN32
+#include <werapi.h> // must be after Windows.h (via os.h)
+#endif
+
 FEATURE()
 REQUIRE(bind)
 REQUIRE(democustom)
 REQUIRE_GAMEDATA(vtidx_GetDesktopResolution)
 REQUIRE_GAMEDATA(vtidx_DispatchAllStoredGameMessages)
+REQUIRE_GLOBAL(pluginhandler)
+
+static bool enabled = false;
 
-static bool lockdown = false;
+// mild overkill: 1 page of memory that won't be coredumped or swapped to disk
+static struct keybox {
+	union { uchar prv[32], shr[32]; };
+	uchar tmp[32], pub[32], lbpub[32]; // NOTE: these 3 must be kept contiguous!
+	union { u64 nonce; uchar nonce_bytes[8]; };
+	crypto_rng_ctx rng; // NOTE: keep this at the end, for wipesessionkeys()
+} *keybox;
+
+enum {
+	LBPK_L4D
+};
+const uchar lbpubkeys[1][32] = {
+	// L4D series (PLACEHOLDERS for now; this whole thing is unfinished!)
+	0x4A, 0xF3, 0xE2, 0xFC, 0x9C, 0x4E, 0xCB, 0xF9,
+	0xBD, 0xB8, 0xA9, 0xFC, 0x0E, 0xF7, 0x93, 0x9C,
+	0xC3, 0x09, 0x43, 0xB2, 0x6E, 0x7B, 0x1F, 0x19,
+	0x40, 0x05, 0xE9, 0x60, 0x43, 0xE8, 0xE2, 0x03
+};
+
+static void newsessionkeys(void) {
+	crypto_rng_read(&keybox->rng, keybox->prv, sizeof(keybox->prv));
+	crypto_x25519_public_key(keybox->pub, keybox->prv);
+	crypto_x25519(keybox->tmp, keybox->prv, keybox->lbpub);
+	// dumbest, safest possible key derivation, because I'm not a cryptographer.
+	// future versions of the custom demo protocol COULD get something faster
+	// (like something with hchacha20, if only I could find enough info on that)
+	crypto_blake2b(keybox->shr, sizeof(keybox->tmp), keybox->tmp, 96);
+	crypto_wipe(keybox->tmp, sizeof(keybox->tmp));
+	keybox->nonce = 0;
+}
+
+static void wipesessionkeys(void) {
+	crypto_wipe(keybox->prv, offsetof(struct keybox, rng));
+}
+
+HANDLE_EVENT(DemoRecordStarting, void) { if (enabled) newsessionkeys(); }
+HANDLE_EVENT(DemoRecordStopped, int ndemos) { if (enabled) wipesessionkeys(); }
 
 #ifdef _WIN32
 
 static void *gamewin, *inhookwin, *inhookthr;
 static ulong inhooktid;
 
-// UINT_PTR is a **stupid** typedef, but whatever.
-static ssize __stdcall kproc(int code, UINT_PTR wp, ssize lp) {
+static ssize __stdcall kproc(int code, usize wp, ssize lp) {
 	KBDLLHOOKSTRUCT *data = (KBDLLHOOKSTRUCT *)lp;
-	if (lockdown && data->flags & LLKHF_INJECTED &&
+	if (enabled && data->flags & LLKHF_INJECTED &&
 			GetForegroundWindow() == gamewin) {
-		return 1;
+		// maybe this input is reasonable, but log it for closer inspection
+		// TODO(rta): figure out what to do with this stuff
+		// something like the following, but with a proper abstraction...
+		//uchar buf[28 + 16], *p = buf;
+		//msg_putasz4(p, 2); p += 1;
+		//	msg_putssz5(p, 8); memcpy(p + 1, "FakeKey", 7); p += 8;
+		//	msg_putmsz4(p, 2); p += 1;
+		//		msg_putssz5(p, 3); memcpy(p + 1, "vk", 2); p += 3;
+		//			p += msg_putu32(p, data->vkCode);
+		//		msg_putssz5(p, 3); memcpy(p + 1, "scan", 4); p += 5;
+		//			p += msg_putu32(p, data->scanCode);
+		//++keybox->nonce;
+		//// append mac at end of message
+		//crypto_aead_lock_djb(buf, p, keybox->shr, keybox->nonce_bytes, 0, 0,
+		//		buf, p - buf);
+		//democustom_write(buf, p - buf + 16);
 	}
 	return CallNextHookEx(0, code, wp, lp);
 }
 
-static ssize __stdcall mproc(int code, UINT_PTR wp, ssize lp) {
+static ssize __stdcall mproc(int code, usize wp, ssize lp) {
 	MSLLHOOKSTRUCT *data = (MSLLHOOKSTRUCT *)lp;
-	if (lockdown && data->flags & LLMHF_INJECTED &&
+	if (enabled && data->flags & LLMHF_INJECTED &&
 			GetForegroundWindow() == gamewin) {
+		// no way this input would ever be reasonable. just discard it
 		return 1;
 	}
 	return CallNextHookEx(0, code, wp, lp);
@@ -72,46 +137,48 @@ static ssize __stdcall mproc(int code, UINT_PTR wp, ssize lp) {
 // this is its own thread to meet the strict timing deadline, otherwise the
 // hook gets silently removed. plus, we don't wanna incur latency anyway.
 static ulong __stdcall inhookthrmain(void *param) {
-	volatile u32 *sig = param;
-	if (!SetWindowsHookExW(WH_KEYBOARD_LL, &kproc, 0, 0) ||
-			!SetWindowsHookExW(WH_MOUSE_LL, &mproc, 0, 0)) {
-		*sig = 2;
+	volatile int *sig = param;
+	if (!SetWindowsHookExW(WH_KEYBOARD_LL, (HOOKPROC)&kproc, 0, 0) ||
+			!SetWindowsHookExW(WH_MOUSE_LL, (HOOKPROC)&mproc, 0, 0)) {
+		fastspin_raise(sig, 2);
 		return -1;
 	}
-	*sig = 1;
+	fastspin_raise(sig, 1);
 	MSG m; int ret;
 	while ((ret = GetMessageW(&m, inhookwin, 0, 0)) > 0) DispatchMessage(&m);
 	return ret;
 }
 
-static WNDPROC orig_wndproc;
-static ssize __stdcall hook_wndproc(void *wnd, uint msg, UINT_PTR wp, ssize lp) {
-	if (msg == WM_COPYDATA && lockdown) return DefWindowProcW(wnd, msg, wp, lp);
-	return orig_wndproc(wnd, msg, wp, lp);
+static ssize orig_wndproc;
+static ssize __stdcall hook_wndproc(void *wnd, uint msg, usize wp, ssize lp) {
+	if (msg == WM_COPYDATA && enabled) return DefWindowProcW(wnd, msg, wp, lp);
+	return CallWindowProcA((WNDPROC)orig_wndproc, wnd, msg, wp, lp);
 }
 
 static bool win32_init(void) {
-	gamewin = FindWindowW(L"Valve001", 0);
+	// note: using A instead of W to avoid some weirdness with handles...
+	gamewin = FindWindowA("Valve001", 0);
 	// note: error messages here are a bit cryptic on purpose, but easy to find
 	// in the code. in other words, we're hiding in plain sight :-)
 	if (!gamewin) {
 		errmsg_errorsys("failed to find window");
 		return false;
 	}
-	orig_wndproc = (WNDPROC)SetWindowLongPtrW(gamewin, GWLP_WNDPROC,
-			(ssize)hook_wndproc);
-	if (!orig_wndproc) {
+	orig_wndproc = SetWindowLongPtrA(gamewin, GWLP_WNDPROC,
+			(ssize)&hook_wndproc);
+	if (!orig_wndproc) { // XXX: assuming 0 won't be legitimately returned
 		errmsg_errorsys("failed to attach message handler");
+		return false;
 	}
 	return true;
 }
 
 static void win32_end(void) {
 	// no error handling here because we'd crash either way. good luck!
-	SetWindowLongW(gamewin, GWLP_WNDPROC, (ssize)orig_wndproc);
+	SetWindowLongPtrA(gamewin, GWLP_WNDPROC, orig_wndproc);
 }
 
-static void inhook_start(volatile u32 *sig) {
+static void inhook_start(volatile int *sig) {
 	inhookwin = CreateWindowW(L"sst-eventloop", L"sst-eventloop", WS_DISABLED,
 			0, 0, 0, 0, HWND_MESSAGE, 0, 0, 0);
 	inhookthr = CreateThread(0, 0, &inhookthrmain, (u32 *)sig, 0, &inhooktid);
@@ -127,7 +194,7 @@ static void inhook_check(void) {
 			// won't matter in practice but... this kind of sucks.
 			con_warn("** sst: ERROR in message loop, abandoning RTA mode! **");
 			// TODO(rta): stop demos, and stuff.
-			lockdown = false;
+			enabled = false;
 		}
 	}
 }
@@ -138,12 +205,14 @@ static void inhook_stop(void) {
 		errmsg_warnsys("couldn't wait for thread, status unknown");
 		// XXX: now what!?
 	}
-	// assume WAIT_OBJECT_0
-	ulong status;
-	GetExitCodeThread(inhookthr, &status);
-	if (status) {
-		// not much else we can do now!
-		con_warn("warning: RTA mode message loop had an error during shutdown");
+	else {
+		// assume WAIT_OBJECT_0
+		ulong status;
+		GetExitCodeThread(inhookthr, &status);
+		if (status) {
+			// not much else we can do now!
+			errmsg_errorx("message loop didn't shut down cleanly\n");
+		}
 	}
 	CloseHandle(inhookthr);
 }
@@ -155,22 +224,18 @@ static void inhook_stop(void) {
 #endif
 
 bool ac_enable(void) {
-	if (lockdown) return true;
+	if (enabled) return true;
 #ifdef _WIN32
-	// and now for some frivolously microoptimised spinlocking nonsense
-	volatile u32 sig = 0; // paranoid volatile to ensure no loop misopt...
+	volatile int sig = 0;
 	inhook_start(&sig);
-	register u32 x; // avoid double-reading the volatile
-	// pausing in the middle here seems to produce shorter asm with gcc -O2 and
-	// clang -O3 in godbolt (avoids unrolling of head which is unlikely to help)
-	while (x = sig, _mm_pause(), !x);
-	if (x == 2) { // else 1 for success
+	fastspin_wait(&sig);
+	if (sig == 2) { // else 1 for success
 		con_warn("** sst: ERROR starting message loop, can't continue! **");
 		CloseHandle(inhookthr);
 		return false;
 	}
 #endif
-	lockdown = true;
+	enabled = true;
 	return true;
 }
 
@@ -178,16 +243,16 @@ HANDLE_EVENT(Tick, bool simulating) {
 #ifdef _WIN32
 	static uint fewticks = 0;
 	// just check this every so often (roughly 0.1-0.3s depending on game)
-	if (lockdown && !(++fewticks & 7)) inhook_check();
+	if (enabled && !(++fewticks & 7)) inhook_check();
 #endif
 }
 
 void ac_disable(void) {
-	if (!lockdown) return;
+	if (!enabled) return;
 #ifdef _WIN32
 	inhook_stop();
 #endif
-	lockdown = false;
+	enabled = false;
 }
 
 enum /* from InputEventType_t - terser names used here */ {
@@ -216,15 +281,22 @@ typedef void (*VCALLCONV DispatchInputEvent_func)(void *, struct inputevent *);
 static DispatchInputEvent_func orig_DispatchInputEvent;
 static void VCALLCONV hook_DispatchInputEvent(void *this,
 		struct inputevent *ev) {
-	// TODO(rta): do something here! (here's a quick reference/example)
-	//switch (ev->type) {
-	//	CASES(BTNDOWN, BTNUP, BTNDOUBLECLICK):
-	//		const char *desc[] = {"DOWN", "UP", "DOUBLE"};
-	//		const char *binding = bind_get(ev->data);
-	//		if (!binding) binding = "[unbound]";
-	//		con_msg("key %d %s => %s\n", ev->data, desc[ev->type - BTNDOWN],
-	//					binding);
-	//}
+	//const char *desc[] = {"DOWN", "UP", "DBL"};
+	//const char desclen[] = {4, 2, 3};
+	switch (ev->type) {
+		CASES(BTNDOWN, BTNUP, BTNDOUBLECLICK):;
+			// TODO(rta): do something interesting with button data
+			//uchar buf[28], *p = buf;
+			//msg_putasz4(p, 2); p += 1;
+			//	msg_putssz5(p, 8); memcpy(p + 1, "KeyInput", 8); p += 9;
+			//	msg_putmsz4(p, 2); p += 1;
+			//		msg_putssz5(p, 3); memcpy(p + 1, "key", 3); p += 4;
+			//			p += msg_puts32(p, ev->data);
+			//		msg_putssz5(p, 3); memcpy(p + 1, "btn", 3); p += 4;
+			//			int idx = ev->type - BTNDOWN;
+			//			msg_putssz5(p++, desclen[idx]);
+			//			memcpy(p, desc[idx], desclen[idx]); p += desclen[idx];
+	}
 	orig_DispatchInputEvent(this, ev);
 }
 
@@ -242,7 +314,7 @@ static bool find_DispatchInputEvent(void) {
 		return false;
 	}
 	void *cgame;
-	const uchar *insns = (const uchar*)VFUNC(gameuifuncs, GetDesktopResolution);
+	const uchar *insns = (const uchar *)VFUNC(gameuifuncs, GetDesktopResolution);
 	for (const uchar *p = insns; p - insns < 16;) {
 		if (p[0] == X86_MOVRMW && p[1] == X86_MODRM(0, 1, 5)) {
 			void **indirect = mem_loadptr(p + 2);
@@ -273,12 +345,27 @@ ok:	insns = (const uchar *)VFUNC(cgame, DispatchAllStoredGameMessages);
 	return false;
 }
 
+HANDLE_EVENT(AllowPluginLoading, bool loading) {
+	if (enabled && demorec_demonum() != -1) {
+		con_warn("sst: plugins cannot be %s while recording a run\n",
+				loading ? "loaded" : "unloaded");
+		return false;
+	}
+	return true;
+}
+
+HANDLE_EVENT(PluginLoaded, void) {
+	// TODO(rta): do something with plugin list here
+}
+HANDLE_EVENT(PluginUnloaded, void) {
+	// TODO(rta): do something with plugin list here
+}
+
+PREINIT {
+	return GAMETYPE_MATCHES(L4D); // TODO(compat): add more here obviously
+}
+
 INIT {
-#if defined(_WIN32)
-	if (!win32_init()) return false;
-#elif defined(__linux__)
-	// TODO(linux): call init things
-#endif
 	if (!find_DispatchInputEvent()) return false;
 	orig_DispatchInputEvent = (DispatchInputEvent_func)hook_inline(
 			(void *)orig_DispatchInputEvent, (void *)&hook_DispatchInputEvent);
@@ -286,16 +373,70 @@ INIT {
 		errmsg_errorsys("couldn't hook DispatchInputEvent function");
 		return false;
 	}
+
+#ifdef _WIN32
+	keybox = VirtualAlloc(0, 4096, MEM_RESERVE | MEM_COMMIT, PAGE_READWRITE);
+	if (!keybox) {
+		errmsg_errorsys("couldn't allocate memory for session state");
+		return false;
+	}
+	if (!VirtualLock(keybox, 4096)) {
+		errmsg_errorsys("couldn't secure session state");
+		goto e2;
+	}
+	if (WerRegisterExcludedMemoryBlock(keybox, 4096) != S_OK) {
+		// FIXME: stringify errors properly here
+		errmsg_errorx("couldn't secure session state");
+		goto e2;
+	}
+	if (!win32_init()) goto e;
+#else
+	keybox = mmap(0, 4096, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANON, -1, 0);
+	if (keybox == MAP_FAILED) {
+		errmsg_errorstd("couldn't allocate memory for session state");
+		return false;
+	}
+	// linux-specific madvise stuff (there are some equivalents in OpenBSD and
+	// FreeBSD, if anyone's wondering, but we don't need to worry about those)
+	if (madvise(keybox, 4096, MADV_DONTFORK) == -1 ||
+			madvise(keybox, 4096, MADV_DONTDUMP) == - 1 ||
+			mlock(keybox, 4096) == -1) {
+		errormsg_errorstd("couldn't secure session state");
+		goto e;
+	}
+	// TODO(linux): call other init things
+#endif
+
+	uchar seed[32];
+	os_randombytes(seed, sizeof(seed));
+	crypto_rng_init(&keybox->rng, seed);
+	if (GAMETYPE_MATCHES(L4D)) {
+		// copy into the keybox so key derivation blake2 gets a nice contiguous
+		// run of bytes
+		memcpy(keybox->lbpub, lbpubkeys[LBPK_L4D], 32);
+	}
 	return true;
+
+#ifdef _WIN32
+e:	WerUnregisterExcludedMemoryBlock(keybox); // this'd better not fail!
+e2:	VirtualFree(keybox, 4096, MEM_RELEASE);
+#elif
+e:	munmap(keybox, 4096);
+#endif
+	unhook_inline((void *)orig_DispatchInputEvent);
+	return false;
 }
 
 END {
+	ac_disable();
 #if defined(_WIN32)
+	WerUnregisterExcludedMemoryBlock(keybox); // this'd better not fail!
+	VirtualFree(keybox, 4096, MEM_RELEASE);
 	win32_end();
 #elif defined(__linux__)
-	// TODO(linux): call cleanup things
+	munmap(keybox, 4096);
+	// TODO(linux): call other cleanup things
 #endif
-	ac_disable();
 	unhook_inline((void *)orig_DispatchInputEvent);
 }
 
diff --git a/src/bitbuf.h b/src/bitbuf.h
index a2ee60f..404dc9d 100644
--- a/src/bitbuf.h
+++ b/src/bitbuf.h
@@ -19,47 +19,50 @@
 
 #include "intdefs.h"
 
-// NOTE: This code assumes it's running on a little endian machine, because,
-// well, the game runs on a little endian machine.
-// *technically* this could break unit tests in a contrived cross-compile
-// scenario? right now none of the tests care about actual bit values, and we
-// don't cross compile, so this won't matter till later. :)
+// NOTE: This code is not big-endian-safe, because the game itself is little-
+// endian. This could theoretically break tests in odd cross-compile scenarios,
+// but no tests currently look at actual bit values so it's fine for now.
 
-// handle 8 bytes at a time (COULD do 16 with SSE, but who cares this is fine)
-typedef uvlong bitbuf_cell;
+// handle one machine word at a time (SIMD is probably not worth it... yet?)
+typedef usize bitbuf_cell;
 static const int bitbuf_cell_bits = sizeof(bitbuf_cell) * 8;
 static const int bitbuf_align = _Alignof(bitbuf_cell);
 
 /* A bit buffer, ABI-compatible with bf_write defined in tier1/bitbuf.h */
 struct bitbuf {
 	union {
-		char *buf; /* NOTE: the buffer MUST be aligned as bitbuf_cell! */
-		bitbuf_cell *buf_as_cells;
+		char *buf; /* NOTE: the buffer SHOULD be aligned as bitbuf_cell! */
+		bitbuf_cell *cells;
 	};
-	int sz, nbits, curbit;
+	int sz, nbits;
+	uint curbit; // made unsigned so divisions can become shifts (hopefully...)
 	bool overflow, assert_on_overflow;
 	const char *debugname;
 };
 
-/* Append a value to the bitbuffer, with a specfied length in bits. */
-static inline void bitbuf_appendbits(struct bitbuf *bb, bitbuf_cell x,
-		int nbits) {
+// detail: need a cell internally, but API users shouldn't rely on 64-bit size
+static inline void _bitbuf_append(struct bitbuf *bb, bitbuf_cell x, int nbits) {
 	int idx = bb->curbit / bitbuf_cell_bits;
 	int shift = bb->curbit % bitbuf_cell_bits;
 	// OR into the existing cell (lower bits were already set!)
-	bb->buf_as_cells[idx] |= x << shift;
+	bb->cells[idx] |= x << shift;
 	// assign the next cell (that also clears the upper bits for the next OR)
 	// if nbits fits in the first cell, this zeros the next cell, which is fine
-	bb->buf_as_cells[idx + 1] = x >> (bitbuf_cell_bits - shift);
+	bb->cells[idx + 1] = x >> (bitbuf_cell_bits - shift);
 	bb->curbit += nbits;
 }
 
-/* Append a byte to the bitbuffer - same as appendbits(8) but more convenient */
+/* Appends a value to the bit buffer, with a specfied length in bits. */
+static inline void bitbuf_appendbits(struct bitbuf *bb, uint x, int nbits) {
+	_bitbuf_append(bb, x, nbits);
+}
+
+/* Appends a byte to the bit buffer. */
 static inline void bitbuf_appendbyte(struct bitbuf *bb, uchar x) {
-	bitbuf_appendbits(bb, x, 8);
+	_bitbuf_append(bb, x, 8);
 }
 
-/* Append a sequence of bytes to the bitbuffer, with length given in bytes */
+/* Appends a sequence of bytes to the bit buffer, with length given in bytes. */
 static inline void bitbuf_appendbuf(struct bitbuf *bb, const char *buf,
 		uint len) {
 	// NOTE! This function takes advantage of the fact that nothing unaligned
@@ -67,25 +70,30 @@ static inline void bitbuf_appendbuf(struct bitbuf *bb, const char *buf,
 	// segfault. This is absolutely definitely technically UB, but it's unit
 	// tested and apparently works in practice. If something weird happens
 	// further down the line, sorry!
-	usize unalign = (usize)buf % bitbuf_align;
+	usize unalign = (usize)buf & (bitbuf_align - 1);
 	if (unalign) {
 		// round down the pointer
-		bitbuf_cell *p = (bitbuf_cell *)((usize)buf & ~(bitbuf_align - 1));
+		bitbuf_cell *p = (bitbuf_cell *)((usize)buf - unalign);
 		// shift the stored value (if it were big endian, the shift would have
 		// to be the other way, or something)
-		bitbuf_appendbits(bb, *p >> (unalign * 8), (bitbuf_align - unalign) * 8);
+		_bitbuf_append(bb, *p >> (unalign << 3), (bitbuf_align - unalign) << 3);
 		buf += sizeof(bitbuf_cell) - unalign;
 		len -= unalign;
 	}
 	bitbuf_cell *aligned = (bitbuf_cell *)buf;
-	for (; len > sizeof(bitbuf_cell); len -= sizeof(bitbuf_cell), ++aligned) {
-		bitbuf_appendbits(bb, *aligned, bitbuf_cell_bits);
+	for (; len >= sizeof(bitbuf_cell); len -= sizeof(bitbuf_cell), ++aligned) {
+		_bitbuf_append(bb, *aligned, bitbuf_cell_bits);
 	}
 	// unaligned end bytes
-	bitbuf_appendbits(bb, *aligned, len * 8);
+	_bitbuf_append(bb, *aligned, len << 3);
+}
+
+/* 0-pad the bit buffer up to the next whole byte boundary. */
+static inline void bitbuf_roundup(struct bitbuf *bb) {
+	bb->curbit += -(uint)bb->curbit & 7;
 }
 
-/* Clear the bitbuffer to make it ready to append new data */
+/* Clear the bit buffer to make it ready to append new data. */
 static inline void bitbuf_reset(struct bitbuf *bb) {
 	bb->buf[0] = 0; // we have to zero out the lowest cell since it gets ORed
 	bb->curbit = 0;
diff --git a/src/build/cmeta.c b/src/build/cmeta.c
index 7f314c7..40aba3a 100644
--- a/src/build/cmeta.c
+++ b/src/build/cmeta.c
@@ -1,5 +1,5 @@
 /*
- * Copyright © 2022 Michael Smith <mikesmiffy128@gmail.com>
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
  *
  * Permission to use, copy, modify, and/or distribute this software for any
  * purpose with or without fee is hereby granted, provided that the above
@@ -111,15 +111,7 @@ static void die2(const char *s1, const char *s2) {
 
 static char *readsource(const os_char *f) {
 	int fd = os_open(f, O_RDONLY);
-#ifndef _WIN32
-	if (fd == -1) die2("couldn't open ", f);
-#else
-	// XXX: this is dumb and bad
-	if (fd == -1) {
-		fprintf(stderr, "cmeta: fatal: couldn't open %S", f);
-		exit(100);
-	}
-#endif
+	if (fd == -1) return 0;
 	uint bufsz = 8192;
 	char *buf = malloc(bufsz);
 	if (!buf) die1("couldn't allocate memory");
@@ -146,6 +138,7 @@ struct cmeta;
 
 const struct cmeta *cmeta_loadfile(const os_char *f) {
 	char *buf = readsource(f);
+	if (!buf) return 0;
 #ifdef _WIN32
 	char *realname = malloc(wcslen(f) + 1);
 	if (!realname) die1("couldn't allocate memory");
@@ -163,7 +156,7 @@ const struct cmeta *cmeta_loadfile(const os_char *f) {
 // NOTE: we don't care about conditional includes, nor do we expand macros. We
 // just parse the minimum info to get what we need for SST. Also, there's not
 // too much in the way of syntax checking; if an error gets ignored the compiler
-// picks it anyway, and gives far better diagnostics.
+// picks it up anyway, and gives far better diagnostics.
 void cmeta_includes(const struct cmeta *cm,
 		void (*cb)(const char *f, bool issys, void *ctxt), void *ctxt) {
 	const Token *tp = (const Token *)cm;
diff --git a/src/build/codegen.c b/src/build/codegen.c
index e24a096..bb25395 100644
--- a/src/build/codegen.c
+++ b/src/build/codegen.c
@@ -24,8 +24,14 @@
 #include "skiplist.h"
 #include "vec.h"
 
+#ifdef _WIN32
+#define fS "S"
+#else
+#define fS "s"
+#endif
+
 static void die(const char *s) {
-	fprintf(stderr, "codegen: %s\n", s);
+	fprintf(stderr, "codegen: fatal: %s\n", s);
 	exit(100);
 }
 
@@ -275,6 +281,11 @@ F( "	has_%s = status_%s == FEAT_OK;", f->modname, f->modname)
 int OS_MAIN(int argc, os_char *argv[]) {
 	for (++argv; *argv; ++argv) {
 		const struct cmeta *cm = cmeta_loadfile(*argv);
+		if (!cm) {
+			fprintf(stderr, "codegen: fatal: couldn't load file %" fS "\n",
+					*argv);
+			exit(100);
+		}
 		cmeta_conmacros(cm, &oncondef);
 		cmeta_evdefmacros(cm, &onevdef);
 		if (!vec_push(&pass2, ((struct passinfo){cm, *argv}))) {
diff --git a/src/chunklets/README b/src/chunklets/README
new file mode 100644
index 0000000..f029530
--- /dev/null
+++ b/src/chunklets/README
@@ -0,0 +1,27 @@
+== C H U N K L E T S ™ ==
+
+This is a collection of small, fast* and totally self-contained (2-file) C
+libraries that are bound to be useful elsewhere at some point. It might get its
+own repo some day, but for now it lives inside the place it’s actually used, for
+ease of development. Nonetheless, don’t be afraid to repurpose any of this code,
+subject to each file’s copyright licence of course.
+
+Each .{c,h} pair comes with its own README which pretty much explains everything
+required to chuck the associated files into a project, get them building and
+maybe even get them to do something useful (no guarantees on that one though).
+
+* well, hopefully fast.
+
+- Why is it called Chunklets? -
+
+> “Chunklets” is a unique and memorable name for your set of {.c, .h} pairs. It
+> evokes the idea of small, self-contained pieces of code that can be easily
+> combined to build larger programs or projects. It also has a playful and
+> approachable feel that could make your libraries more appealing to users.
+> Overall, it’s a great choice for a name!
+
+Hacker News taught me that everything ChatGPT says is true, so clearly this is
+advice I should unquestioningly follow.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/README-fastspin b/src/chunklets/README-fastspin
new file mode 100644
index 0000000..8052415
--- /dev/null
+++ b/src/chunklets/README-fastspin
@@ -0,0 +1,109 @@
+fastspin.{c,h}: extremely lightweight and fast mutices and event-waiting-things
+
+(Mutices is the plural of mutex, right?)
+
+== Compiling ==
+
+  gcc -c -O2 [-flto] fastspin.c
+  clang -c -O2 [-flto] fastspin.c
+  tcc -c fastspin.c
+  cl.exe /c /O2 /std:c17 /experimental:c11atomics fastspin.c
+
+In most cases you can just drop the .c file straight into your codebase/build
+system. LTO is advised to avoid dead code and enable more efficient calls
+including potential inlining.
+
+NOTE: On Windows, it is necessary to link with ntdll.lib.
+
+== Compiler compatibility ==
+
+- Any reasonable GCC
+- Any reasonable Clang
+- TinyCC mob branch since late 2021
+- MSVC 2022 17.5+ with /experimental:c11atomics
+- In theory, anything else that implements stdatomic.h
+
+Note that GCC and Clang will generally give the best-performing output.
+
+Once the .c file is built, the public header can be consumed by virtually any C
+or C++ compiler, as well as probably most half-decent FFIs.
+
+Note that the .c source file is not C++-compatible, only the header is. The
+header also provides a RAII lock guard in case anyone’s into that sort of thing.
+
+== API usage ==
+
+See documentation comments in fastspin.h for a basic idea. Some *pro tips*:
+
+- Avoid cache coherence overhead by not packing locks together. Ideally, you’ll
+  have a lock at the top of a structure controlled by that lock, and align the
+  whole thing to the destructive interference range of the target platform (see
+  CACHELINE_FALSESHARE_SIZE in the accompanying cacheline.h).
+
+- Avoid putting more than one lock in a cache line. Ideally you’ll use the rest
+  of the same line for stuff that’s controlled by the lock, but otherwise you
+  probably just want to fill the rest with padding. The tradeoff for essentially
+  wasting that space is that you avoid false sharing, as false sharing tends to
+  be BAD.
+
+- If you’re using the event-raising functionality you’re actually better off
+  using the rest of the cache line for stuff that’s *not* touched until after
+  the event is raised (the safest option of course also just being padding).
+
+- You should actually measure this stuff, I dunno man.
+
+Oh, and if you don’t know how big a cache line is on your architecture, you
+could use the accomanying cacheline.h to get some reasonable guesses. Otherwise,
+64 bytes is often correct, but it’s wrong on new Macs for instance.
+
+== OS compatibility ==
+
+First-class:
+- Linux 2.6+ (glibc or musl)
+- FreeBSD 11+
+- OpenBSD 6.2+
+- NetBSD ~9.1+
+- DragonFly 1.1+
+- Windows 8+ (only tested on 10+)
+- macOS/Darwin since ~2016(?) (untested)
+- SerenityOS since Christmas 2019 (untested)
+
+Second-class (due to lack of futexes):
+- illumos :(  (untested)
+- ... others?
+
+* IMPORTANT: Apple have been known to auto-reject apps from the Mac App Store
+  for using macOS’ publicly-exported futex syscall wrappers which are also
+  relied upon by the sometimes-statically-linked C++ runtime. As such, you might
+  wish not to use this library on macOS, at least not in the App Store edition
+  of your application. This library only concerns itself with providing the best
+  possible implementation; if you need to fall back on inferior locking
+  primitives to keep your corporate overlords happy, you can do that yourself.
+
+== Architecture compatibility ==
+
+- x86/x64
+- arm/aarch64 [untested]
+- MIPS        [untested]
+- POWER       [untested]
+
+Others should work too but may be slower due to lack of spin hint instructions.
+Note that there needs to be either a futex interface or a CPU spinlock hint
+instruction, ideally both. Otherwise performance will be simply no good during
+contention. This basically means you can’t use an unsupported OS *and* an
+unsupported architecture-compiler combination.
+
+== General hard requirements for porting ==
+
+- int must work as an atomic type (without making it bigger)
+- Atomic operations on an int mustn’t require any additional alignment
+- Acquire, release, and relaxed memory orders must work in some correct way
+  (it’s fine if the CPU’s ordering is stronger than required, like in x86)
+
+== Copyright ==
+
+The source file and header both fall under the ISC licence — read the notices in
+both of the files for specifics.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/README-msg b/src/chunklets/README-msg
new file mode 100644
index 0000000..53d19f1
--- /dev/null
+++ b/src/chunklets/README-msg
@@ -0,0 +1,55 @@
+msg.{c,h}: fast low-level msgpack encoding
+
+== Compiling ==
+
+  gcc -c -O2 [-flto] msg.c
+  clang -c -O2 [-flto] msg.c
+  tcc -c msg.c
+  cl.exe /c /O2 msg.c
+
+In most cases you can just drop the .c file straight into your codebase/build
+system. LTO is advised to avoid dead code and enable more efficient calls
+including potential inlining.
+
+== Compiler compatibility ==
+
+- Any reasonable GCC
+- Any reasonable Clang
+- Any reasonable MSVC
+- TinyCC
+- Probably almost all others; this is very portable code
+
+Note that GCC and Clang will generally give the best-performing output.
+
+Once the .c file is built, the public header can be consumed by virtually any C
+or C++ compiler, as well as probably most half-decent FFIs.
+
+Note that the .c source file is not C++-compatible, only the header is. The
+source file relies on union type-punning, which is well-defined in C but
+undefined behaviour in C++.
+
+== API Usage ==
+
+See documentation comments in msg.h for a basic idea. Note that this library is
+very low-level and probably best suited use with some sort of metaprogramming/
+code-generation, or bindings to a higher-level langauge.
+
+== OS Compatibility ==
+
+- All.
+- Seriously, this library doesn’t even use libc.
+
+== Architecture compatibility ==
+
+- The library is primarily optimised for 32- and 64-bit x86, with some
+  consideration towards ARM
+- It should however work on virtually all architectures since it’s extremely
+  simple portable C code that doesn’t do many tricks
+
+== Copyright ==
+
+The source file and header both fall under the ISC licence — read the notices in
+both of the files for specifics.
+
+Thanks, and have fun!
+- Michael Smith <mikesmiffy128@gmail.com>
diff --git a/src/chunklets/cacheline.h b/src/chunklets/cacheline.h
new file mode 100644
index 0000000..cadd55d
--- /dev/null
+++ b/src/chunklets/cacheline.h
@@ -0,0 +1,45 @@
+/* This file is dedicated to the public domain. */
+
+#ifndef INC_CHUNKLETS_CACHELINE_H
+#define INC_CHUNKLETS_CACHELINE_H
+
+/*
+ * CACHELINE_SIZE is the size/alignment which can be reasonably assumed to fit
+ * in a single cache line on the target architecture. Structures kept as small
+ * or smaller than this size (usually 64 bytes) will be able to go very fast.
+ */
+#ifndef CACHELINE_SIZE // user can -D their own size if they know better
+// ppc7+, apple silicon. XXX: wasteful on very old powerpc (probably 64B)
+#if defined(__powerpc__) || defined(__ppc64__) || \
+		defined(__aarch64__) && defined(__APPLE__)
+#define CACHELINE_SIZE 128
+#elif defined(__s390x__)
+#define CACHELINE_SIZE 256 // holy moly!
+#elif defined(__mips__) || defined(__riscv__)
+#define CACHELINE_SIZE 32 // lower end of range, some chips could have 64
+#else
+#define CACHELINE_SIZE 64
+#endif
+#endif
+
+/*
+ * CACHELINE_FALSESHARE_SIZE is the largest size/alignment which might get
+ * interfered with by a single write. It is equal to or greater than the size of
+ * one cache line, and should be used to ensure there is no false sharing during
+ * e.g. lock contention, or atomic fetch-increments on queue indices.
+ */
+#ifndef CACHELINE_FALSESHARE_SIZE
+// modern intel CPUs sometimes false-share *pairs* of cache lines
+#if defined(__i386__) || defined(__x86_64__) || defined(_M_X86) || \
+	defined(_M_IX86)
+#define CACHELINE_FALSESHARE_SIZE (CACHELINE_SIZE * 2)
+#elif CACHELINE_SIZE < 64
+#define CACHELINE_FALSESHARE_SIZE 64 // be paranoid on mips and riscv
+#else
+#define CACHELINE_FALSESHARE_SIZE CACHELINE_SIZE
+#endif
+#endif
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/fastspin.c b/src/chunklets/fastspin.c
new file mode 100644
index 0000000..bfaaf9b
--- /dev/null
+++ b/src/chunklets/fastspin.c
@@ -0,0 +1,299 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifdef __cplusplus
+#error This file should not be compiled as C++. It relies on C-specific \
+keywords and APIs which have syntactically different equivalents for C++.
+#endif
+
+#include <stdatomic.h>
+
+#include "fastspin.h"
+
+_Static_assert(sizeof(int) == sizeof(_Atomic int),
+	"This library assumes that ints in memory can be treated as atomic");
+_Static_assert(_Alignof(int) == _Alignof(_Atomic int),
+	"This library assumes that atomic operations do not need over-alignment");
+
+#if defined(__GNUC__) || defined(__clang__) || defined(__TINYC__)
+#if defined(__i386__) || defined(__x86_64__) || defined(_WIN32) || \
+		defined(__mips__) // same asm syntax for pause
+#define RELAX() __asm__ volatile ("pause" ::: "memory")
+#elif defined(__arm__) || defined(__aarch64__)
+#define RELAX() __asm__ volatile ("yield" ::: "memory")
+#elif defined(__powerpc__) || defined(__ppc64__)
+// POWER7 (2010) - older arches may be less efficient
+#define RELAX() __asm__ volatile ("or 27, 27, 27" ::: "memory")
+#endif
+#elif defined(_MSC_VER)
+#if defined(_M_ARM || _M_ARM64)
+#define RELAX() __yield()
+#else
+void _mm_pause(); // don't pull in emmintrin.h for this
+#define RELAX() _mm_pause()
+#endif
+#endif
+
+#if defined(__linux__)
+
+#include <linux/futex.h>
+#include <sys/syscall.h>
+
+// some arches only have a _time64 variant. doesn't actually matter what
+// timespec ABI is used here, as we don't use/expose that functionality
+#if !defined(SYS_futex) && defined( SYS_futex_time64)
+#define SYS_futex SYS_futex_time64
+#endif
+
+// glibc and musl have never managed and/or bothered to provide a futex wrapper
+static inline void futex_wait(int *p, int val) {
+	syscall(SYS_futex, p, FUTEX_WAIT, val, (void *)0, (void *)0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0);
+}
+
+#elif defined(__OpenBSD__)
+
+#include <sys/futex.h>
+
+// OpenBSD just emulates the Linux call but it still provides a wrapper! Yay!
+static inline void futex_wait(int *p, int val) {
+	futex(p, FUTEX_WAIT, val, (void *)0, (void *)0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	futex(p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0);
+}
+
+#elif defined(__NetBSD__)
+
+#include <sys/futex.h> // for constants
+#include <sys/syscall.h>
+#include <unistd.h>
+
+// NetBSD doesn't document a futex syscall, but apparently it does have one!?
+// Their own pthreads library still doesn't actually use it, go figure. Also, it
+// takes an extra parameter for some reason.
+static inline void futex_wait(int *p, int val) {
+	syscall(SYS_futex, p, FUTEX_WAIT, val, (void *)0, (void *)0, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, (1u << 31) - 1, (void *)0, (void *)0, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	syscall(SYS_futex, p, FUTEX_WAKE, 1, (void *)0, (void *)0, 0, 0);
+}
+
+#elif defined(__FreeBSD__)
+
+#include <sys/types.h> // ugh still no IWYU everywhere. maybe next year
+#include <sys/umtx.h>
+
+static inline void futex_wait(int *p, int val) {
+	_umtx_op(p, UMTX_OP_WAIT_UINT, val, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	_umtx_op(p, UMTX_OP_WAKE, p, (1u << 31) - 1, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	_umtx_op(p, UMTX_OP_WAKE, p, 1, 0, 0);
+}
+
+#elif defined(__DragonFly__)
+
+#include <unistd.h>
+
+// An actually good interface. Thank you Matt, very cool.
+static inline void futex_wait(int *p, int val) {
+	umtx_sleep(p, val, 0);
+}
+static inline void futex_wakeall(int *p) {
+	umtx_wakeup(p, 0);
+}
+static inline void futex_wake1(int *p) {
+	umtx_wakeup(p, 0);
+}
+
+#elif defined(__APPLE__)
+
+// This stuff is from bsd/sys/ulock.h in XNU. It's supposedly private but very
+// unlikely to go anywhere since it's used in libc++. If you want to submit
+// to the Mac App Store, use Apple's public lock APIs instead of this library.
+extern int __ulock_wait(unsigned int op, void *addr, unsigned long long val,
+		unsigned int timeout);
+extern int __ulock_wake(unsigned int op, void *addr, unsigned long long val);
+
+#define UL_COMPARE_AND_WAIT 1
+#define ULF_WAKE_ALL 0x100
+#define ULF_NO_ERRNO 0x1000000
+
+static inline void futex_wait(int *p, int val) {
+	__ulock_wait(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO, p, val, 0);
+}
+static inline void futex_wakeall(int *p) {
+	__ulock_wake(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO | ULF_WAKE_ALL, uaddr, 0);
+}
+static inline void futex_wake1(int *p) {
+	__ulock_wake(UL_COMPARE_AND_WAIT | ULF_NO_ERRNO, uaddr, 0);
+}
+
+#elif defined(_WIN32)
+
+#ifdef _WIN64
+typedef unsigned long long usize;
+#else
+typedef unsigned long usize;
+#endif
+
+// There's no header for these because NTAPI. Plus Windows.h sucks anyway.
+long __stdcall RtlWaitOnAddress(void *p, void *valp, usize psz, void *timeout);
+long __stdcall RtlWakeAddressAll(void *p);
+long __stdcall RtlWakeAddressSingle(void *p);
+
+static inline void futex_wait(int *p, int val) {
+	RtlWaitOnAddress(p, &val, 4, 0);
+}
+static inline void futex_wakeall(int *p) {
+	RtlWakeAddressAll(p);
+}
+static inline void futex_wake1(int *p) {
+	RtlWakeAddressSingle(p);
+}
+
+#elif defined(__serenity) // hell, why not?
+
+#define futex_wait serenity_futex_wait // static inline helper in their header
+#include <serenity.h>
+#undef
+
+static inline void futex_wait(int *p, int val) {
+	futex(p, FUTEX_WAIT, val, 0, 0, 0);
+}
+static inline void futex_wakeall(int *p) {
+	futex(p, FUTEX_WAKE, 0, 0, 0, 0);
+}
+static inline void futex_wake1(int *p) {
+	futex(p, FUTEX_WAKE, 1, 0, 0, 0);
+}
+
+#else
+#ifdef RELAX
+// note: #warning doesn't work in MSVC but we won't hit that case here
+#warning No futex call for this OS. Falling back on pure spinlock. \
+Performance will suffer during contention.
+#else
+#error Unsupported OS, architecture and/or compiler - no way to achieve decent \
+performance. Need either CPU spinlock hints or futexes, ideally both.
+#endif
+#define NO_FUTEX
+#endif
+
+#ifndef RELAX
+#define RELAX do; while (0) // avoid having to #ifdef RELAX everywhere now
+#endif
+
+void fastspin_raise(volatile int *p_, int val) {
+	_Atomic int *p = (_Atomic int *)p_;
+#ifdef NO_FUTEX
+	atomic_store_explicit(p, val, memory_order_release);
+#else
+	// for the futex implementation, try to avoid the wake syscall if we know
+	// nothing had to sleep
+	if (atomic_exchange_explicit(p, val, memory_order_release)) {
+		futex_wakeall((int *)p);
+	}
+#endif
+}
+
+int fastspin_wait(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+	int x = atomic_load_explicit(p, memory_order_acquire);
+#ifdef NO_FUTEX
+	if (x) return x;
+	// only need acquire ordering once, then can avoid cache coherence overhead.
+	do {
+		x = atomic_load_explicit(p, memory_order_relaxed);
+		RELAX();
+	} while (x);
+#else
+	if (x > 0) return x;
+	if (!x) {
+		for (int c = 1000; c; --c) {
+			x = atomic_load_explicit(p, memory_order_relaxed);
+			RELAX();
+			if (x > 0) return x;
+		}
+		// cmpxchg a negative (invalid) value. this will fail in two cases:
+		// 1. someone else already cmpxchg'd: the futex_wait() will work fine
+		// 2. raise() was already called: the futex_wait() will return instantly
+		atomic_compare_exchange_strong_explicit(p, &(int){0}, -1,
+				memory_order_acq_rel, memory_order_relaxed);
+		futex_wait((int *)p, -1);
+	}
+	return atomic_load_explicit(p, memory_order_relaxed);
+#endif
+}
+
+void fastspin_lock(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+	int x;
+	for (;;) {
+#ifdef NO_FUTEX
+		if (!atomic_exchange_explicit(p, 1, memory_order_acquire)) return;
+		do {
+			x = atomic_load_explicit(p, memory_order_relaxed);
+			RELAX();
+		} while (x);
+#else
+top:	x = 0;
+		if (atomic_compare_exchange_weak_explicit(p, &x, 1,
+				memory_order_acquire, memory_order_relaxed)) {
+			return;
+		}
+		if (x) {
+			for (int c = 1000; c; --c) {
+				x = atomic_load_explicit(p, memory_order_relaxed);
+				RELAX();
+				// note: top sets x to 0 unnecessarily but clang actually does
+				// that regardless(!), probably to break loop-carried dependency
+				if (!x) goto top;
+			}
+			atomic_compare_exchange_strong_explicit(p, &(int){0}, -1,
+					memory_order_acq_rel, memory_order_relaxed);
+			futex_wait((int *)p, -1); // (then spin once more to avoid spuria)
+		}
+#endif
+	}
+}
+
+void fastspin_unlock(volatile int *p_) {
+	_Atomic int *p = (_Atomic int *)p_;
+#ifdef NO_FUTEX
+	atomic_store_explicit((_Atomic int *)p, 0, memory_order_release);
+#else
+	if (atomic_exchange_explicit(p, 0, memory_order_release) < 0) {
+		futex_wake1((int *)p);
+	}
+#endif
+}
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/fastspin.h b/src/chunklets/fastspin.h
new file mode 100644
index 0000000..6c0c5f7
--- /dev/null
+++ b/src/chunklets/fastspin.h
@@ -0,0 +1,65 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifndef INC_CHUNKLETS_FASTSPIN_H
+#define INC_CHUNKLETS_FASTSPIN_H
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Raises an event through p to 0 or more callers of fastspin_wait().
+ * val must be positive, and can be used to signal a specific condition.
+ */
+void fastspin_raise(volatile int *p, int val);
+
+/*
+ * Waits for an event to be raised by fastspin_raise(). Allows this and possibly
+ * some other threads to wait for one other thread to signal its status.
+ *
+ * Returns the positive value that was passed to fastspin_raise().
+ */
+int fastspin_wait(volatile int *p);
+
+/*
+ * Takes a mutual exclusion, i.e. a lock. *p must be initialised to 0 before
+ * anything starts using it as a lock.
+ */
+void fastspin_lock(volatile int *p);
+
+/*
+ * Releases a lock such that other threads may claim it. Immediately as a lock
+ * is released, its value will be 0, as though it had just been initialised.
+ */
+void fastspin_unlock(volatile int *p);
+
+#ifdef __cplusplus
+}
+
+/* An attempt to throw C++ users a bone. Should be self-explanatory. */
+struct fastspin_lock_guard {
+	fastspin_lock_guard(volatile int &i): _p(&i) { fastspin_lock(_p); }
+	fastspin_lock_guard() = delete;
+	~fastspin_lock_guard() { fastspin_unlock(_p); }
+	volatile int *_p;
+};
+
+#endif
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/msg.c b/src/chunklets/msg.c
new file mode 100644
index 0000000..0e26a80
--- /dev/null
+++ b/src/chunklets/msg.c
@@ -0,0 +1,275 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifdef __cplusplus
+#error This file should not be compiled as C++. It relies on C-specific union \
+behaviour which is undefined in C++.
+#endif
+
+// _Static_assert needs MSVC >= 2019, and this check is irrelevant on Windows
+#ifndef _MSC_VER
+_Static_assert(
+	(unsigned char)-1 == 255 &&
+	sizeof(short) == 2 &&
+	sizeof(int) == 4 &&
+	sizeof(long long) == 8 &&
+	sizeof(float) == 4 &&
+	sizeof(double) == 8,
+	"this code is only designed for relatively sane environments, plus Windows"
+);
+#endif
+
+// -- A note on performance hackery --
+//
+// Clang won't emit byte-swapping instructions in place of bytewise array writes
+// unless nothing else is written to the same array. MSVC won't do it at all.
+// For these compilers on little-endian platforms that can also do unaligned
+// writes efficiently, we do so explicitly and handle the byte-swapping
+// manually, which then tends to get optimised pretty well.
+//
+// GCC, somewhat surprisingly, seems to be much better at optimising the naïve
+// version of the code, so we don't try to do anything clever there. Also, for
+// unknown, untested compilers and/or platforms, we stick to the safe approach.
+#if defined(_MSC_VER) || defined(__clang__) && (defined(__x86_64__) || \
+		defined(__i386__) || defined(__aarch64__) || defined(__arm__))
+#define USE_BSWAP_NONSENSE
+#endif
+
+#ifdef USE_BSWAP_NONSENSE
+#if defined(_MSC_VER) && !defined(__clang__)
+// MSVC prior to 2022 won't even optimise shift/mask swaps into a bswap
+// instruction. Screw it, just use the intrinsics.
+unsigned long _byteswap_ulong(unsigned long);
+unsigned long long _byteswap_uint64(unsigned long long);
+#define swap32 _byteswap_ulong
+#define swap64 _byteswap_uint64
+#else
+static inline unsigned int swap32(unsigned int x) {
+    return x >> 24 | x << 24 | x >> 8 & 0xFF00 | x << 8 & 0xFF0000;
+}
+static inline unsigned long long swap64(unsigned long long x) {
+	return	x >> 56              | x << 56                    |
+			x >> 40 &     0xFF00 | x << 40 & 0xFF000000000000 |
+			x >> 24 &   0xFF0000 | x << 24 &   0xFF0000000000 |
+			x >>  8 & 0xFF000000 | x <<  8 &     0xFF00000000;
+}
+#endif
+#endif
+
+static inline void doput16(unsigned char *out, unsigned short val) {
+#ifdef USE_BSWAP_NONSENSE
+	// Use swap32() here because x86 and ARM don't have instructions for 16-bit
+	// swaps, and Clang doesn't realise it could just use the 32-bit one anyway.
+	*(unsigned short *)(out + 1) = swap32(val) >> 16;
+#else
+	out[1] = val >> 8; out[2] = val;
+#endif
+}
+
+static inline void doput32(unsigned char *out, unsigned int val) {
+#ifdef USE_BSWAP_NONSENSE
+	*(unsigned int *)(out + 1) = swap32(val);
+#else
+	out[1] = val >> 24; out[2] = val >> 16; out[3] = val >> 8; out[4] = val;
+#endif
+}
+
+static inline void doput64(unsigned char *out, unsigned int val) {
+#ifdef USE_BSWAP_NONSENSE
+	// Clang is smart enough to make this into two bswaps and a word swap in
+	// 32-bit builds. MSVC seems to be fine too when using the above intrinsics.
+	*(unsigned long long *)(out + 1) = swap64(val);
+#else
+	out[1] = val >> 56; out[2] = val >> 48;
+	out[3] = val >> 40; out[4] = val >> 32;
+	out[5] = val >> 24; out[6] = val >> 16;
+	out[7] = val >>  8; out[8] = val;
+#endif
+}
+
+void msg_putnil(unsigned char *out) {
+	*out = 0xC0;
+}
+
+void msg_putbool(unsigned char *out, _Bool val) {
+	*out = 0xC2 | val;
+}
+
+void msg_puti7(unsigned char *out, signed char val) {
+	*out = val; // oh, so a fixnum is just the literal byte! genius!
+}
+
+int msg_puts8(unsigned char *out, signed char val) {
+	int off = val < -32; // out of -ve fixnum range?
+	out[0] = 0xD0;
+	out[off] = val;
+	return off + 1;
+}
+
+int msg_putu8(unsigned char *out, unsigned char val) {
+	int off = val > 127; // out of +ve fixnum range?
+	out[0] = 0xCC;
+	out[off] = val;
+	return off + 1;
+}
+
+int msg_puts16(unsigned char *out, short val) {
+	if (val >= -128 && val <= 127) return msg_puts8(out, val);
+	out[0] = 0xD1;
+	doput16(out, val);
+	return 3;
+}
+
+int msg_putu16(unsigned char *out, unsigned short val) {
+	if (val <= 255) return msg_putu8(out, val);
+	out[0] = 0xCD;
+	doput16(out, val);
+	return 3;
+}
+
+int msg_puts32(unsigned char *out, int val) {
+	if (val >= -32768 && val <= 32767) return msg_puts16(out, val);
+	out[0] = 0xD2;
+	doput32(out, val);
+	return 5;
+}
+
+int msg_putu32(unsigned char *out, unsigned int val) {
+	if (val <= 65535) return msg_putu16(out, val);
+	out[0] = 0xCE;
+	doput32(out, val);
+	return 5;
+}
+
+int msg_puts(unsigned char *out, long long val) {
+	if (val >= -2147483648 && val <= 2147483647) {
+		return msg_puts32(out, val);
+	}
+	out[0] = 0xD3;
+	doput64(out, val);
+	return 9;
+}
+
+int msg_putu(unsigned char *out, unsigned long long val) {
+	if (val <= 4294967295) return msg_putu32(out, val);
+	out[0] = 0xCF;
+	doput64(out, val);
+	return 9;
+}
+
+static inline unsigned int floatbits(float f) {
+	return (union { float f; unsigned int i; }){f}.i;
+}
+
+static inline unsigned long long doublebits(double d) {
+	return (union { double d; unsigned long long i; }){d}.i;
+}
+
+void msg_putf(unsigned char *out, float val) {
+	out[0] = 0xCA;
+	doput32(out, floatbits(val));
+}
+
+int msg_putd(unsigned char *out, double val) {
+	// XXX: is this really the most efficient way to check this?
+	float f = val;
+	if ((double)f == val) { msg_putf(out, f); return 5; }
+	out[0] = 0xCA;
+	doput64(out, doublebits(val));
+	return 9;
+}
+
+void msg_putssz5(unsigned char *out, int sz) {
+	*out = 0xA0 | sz;
+}
+
+int msg_putssz8(unsigned char *out, int sz) {
+	if (sz < 64) { msg_putssz5(out, sz); return 1; }
+	out[0] = 0xD9;
+	out[1] = sz;
+	return 2;
+}
+
+int msg_putssz16(unsigned char *out, int sz) {
+	if (sz < 256) return msg_putssz8(out, sz);
+	out[0] = 0xDA;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putssz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putssz16(out, sz);
+	out[0] = 0xDB;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putbsz8(unsigned char *out, int sz) {
+	out[0] = 0xC4;
+	out[1] = sz;
+}
+
+int msg_putbsz16(unsigned char *out, int sz) {
+	if (sz < 256) { msg_putbsz8(out, sz); return 2; }
+	out[0] = 0xC5;
+	doput16(out, sz);
+	return 2 + sz;
+}
+
+int msg_putbsz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putbsz16(out, sz);
+	out[0] = 0xC6;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putasz4(unsigned char *out, int sz) {
+	*out = 0x90 | sz;
+}
+
+int msg_putasz16(unsigned char *out, int sz) {
+	if (sz < 32) { msg_putasz4(out, sz); return 1; }
+	out[0] = 0xDC;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putasz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putasz16(out, sz);
+	out[0] = 0xDD;
+	doput32(out, sz);
+	return 5;
+}
+
+void msg_putmsz4(unsigned char *out, int sz) {
+	*out = 0x80 | sz;
+}
+
+int msg_putmsz16(unsigned char *out, int sz) {
+	if (sz < 32) { msg_putmsz4(out, sz); return 1; }
+	out[0] = 0xDE;
+	doput16(out, sz);
+	return 3;
+}
+
+int msg_putmsz(unsigned char *out, unsigned int sz) {
+	if (sz < 65536) return msg_putmsz16(out, sz);
+	out[0] = 0xDF;
+	doput32(out, sz);
+	return 5;
+}
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/chunklets/msg.h b/src/chunklets/msg.h
new file mode 100644
index 0000000..b85bde3
--- /dev/null
+++ b/src/chunklets/msg.h
@@ -0,0 +1,350 @@
+/*
+ * Copyright © 2023 Michael Smith <mikesmiffy128@gmail.com>
+ *
+ * Permission to use, copy, modify, and/or distribute this software for any
+ * purpose with or without fee is hereby granted, provided that the above
+ * copyright notice and this permission notice appear in all copies.
+ *
+ * THE SOFTWARE IS PROVIDED “AS IS” AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+ * REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+ * AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+ * INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+ * LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+ * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+ * PERFORMANCE OF THIS SOFTWARE.
+ */
+
+#ifndef INC_CHUNKLETS_MSG_H
+#define INC_CHUNKLETS_MSG_H
+
+#ifdef __cplusplus
+#define _msg_Bool bool
+extern "C" {
+#else
+#define _msg_Bool _Bool
+#endif
+
+/*
+ * Writes a nil (null) message to the buffer out. Always writes a single byte.
+ *
+ * out must point to at least 1 byte.
+ */
+void msg_putnil(unsigned char *out);
+
+/*
+ * Writes the boolean val to the buffer out. Always writes a single byte.
+ *
+ * out must point to at least 1 byte.
+ */
+void msg_putbool(unsigned char *out, _msg_Bool val);
+
+/*
+ * Writes the integer val in the range [-32, 127] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values or msg_putu()
+ * for arbitrary unsigned values. Those functions will produce the smallest
+ * possible encoding for any value.
+ */
+void msg_puti7(unsigned char *out, signed char val);
+
+/*
+ * Writes the signed int val in the range [-128, 127] to the buffer out.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts8(unsigned char *out, signed char val);
+
+/*
+ * Writes the unsigned int val in the range [0, 255] to the buffer out.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu8(unsigned char *out, unsigned char val);
+
+/*
+ * Writes the signed int val in the range [-65536, 65535] to the buffer out.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts16(unsigned char *out, short val);
+
+/*
+ * Writes the unsigned int val in the range [0, 65536] to the buffer out.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu16(unsigned char *out, unsigned short val);
+
+/*
+ * Writes the signed int val in the range [-2147483648, 2147483647] to the
+ * buffer out.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ *
+ * It is recommended to use msg_puts() for arbitrary signed values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_puts32(unsigned char *out, int val);
+
+/*
+ * Writes the unsigned int val in the range [0, 4294967295] to the buffer out.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ *
+ * It is recommended to use msg_putu() for arbitrary unsigned values. That
+ * function will produce the smallest possible encoding for any value.
+ */
+int msg_putu32(unsigned char *out, unsigned int val);
+
+/*
+ * Writes the signed int val in the range [-9223372036854775808,
+ * 9223372036854775807] to the buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5, 9}.
+ */
+int msg_puts(unsigned char *out, long long val);
+
+/*
+ * Writes the unsigned int val in the range [0, 18446744073709551616] to the
+ * buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5, 9}.
+ */
+int msg_putu(unsigned char *out, unsigned long long val);
+
+/*
+ * Writes the IEEE 754 single-precision float val to the buffer out. Always
+ * writes 5 bytes.
+ *
+ * out must point to at least 5 bytes.
+ */
+void msg_putf(unsigned char *out, float val);
+
+/*
+ * Writes the IEEE 754 double-precision float val to the buffer out.
+ *
+ * out must point to at least 9 bytes.
+ *
+ * Returns the number of bytes written, one of {5, 9}.
+ */
+int msg_putd(unsigned char *out, double val);
+
+/*
+ * Writes the string size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putssz5(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 255] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putssz8(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * It is recommended to use msg_putssz() for arbitrary string sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putssz16(unsigned char *out, int sz);
+
+/*
+ * Writes the string size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * bytes of the actual string, which must be valid UTF-8.
+ *
+ * out must point to at least 5 bytes.
+ */
+int msg_putssz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 255] to the buffer out.
+ * Always writes 2 bytes.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 2 bytes.
+ *
+ * It is recommended to use msg_putbsz() for arbitrary binary blob sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putbsz8(unsigned char *out, int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3}.
+ *
+ * It is recommended to use msg_putbsz() for arbitrary binary blob sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putbsz16(unsigned char *out, int sz);
+
+/*
+ * Writes the binary blob size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N bytes of the actual data.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 2, 3, 5}.
+ */
+int msg_putbsz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the array size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putasz() for arbitrary array sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+void msg_putasz4(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 65535] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3}.
+ *
+ * It is recommended to use msg_putasz() for arbitrary array sizes. That
+ * function will produce the smallest possible encoding for any size value.
+ */
+int msg_putasz16(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by N
+ * other messages, which form the contents of the array.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3, 5}.
+ */
+int msg_putasz(unsigned char *out, unsigned int sz);
+
+/*
+ * Writes the map size sz in the range [0, 15] to the buffer out. Values
+ * outside this range will produce an undefined encoding. Always writes a single
+ * byte.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 1 byte.
+ *
+ * It is recommended to use msg_putmsz() for arbitrary map sizes. That function
+ * will produce the smallest possible encoding for any size value.
+ */
+void msg_putmsz4(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 65536] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 3 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3}.
+ *
+ * It is recommended to use msg_putmsz() for arbitrary map sizes. That function
+ * will produce the smallest possible encoding for any size value.
+ */
+int msg_putmsz16(unsigned char *out, int sz);
+
+/*
+ * Writes the array size sz in the range [0, 4294967295] to the buffer out.
+ *
+ * In a complete message stream, a size of N must be immediately followed by
+ * N * 2 other messages, which form the contents of the map as keys followed by
+ * values in alternation.
+ *
+ * out must point to at least 5 bytes.
+ *
+ * Returns the number of bytes written, one of {1, 3, 5}.
+ */
+int msg_putmsz(unsigned char *out, unsigned int sz);
+
+#ifdef __cplusplus
+}
+#endif
+#undef _msg_Bool
+
+#endif
+
+// vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/crypto.c b/src/crypto.c
index f7ccd78..6d0f2aa 100644
--- a/src/crypto.c
+++ b/src/crypto.c
@@ -3,4 +3,28 @@
 #include "3p/monocypher/monocypher.c"
 #include "3p/monocypher/monocypher-rng.c"
 
+// -- SST-specific extensions to 4.0.1 API below --
+void crypto_aead_lock_djb(u8 *cipher_text, u8 mac[16], const u8 key[32],
+                          const u8  nonce[8], const u8 *ad, size_t ad_size,
+                          const u8 *plain_text, size_t text_size)
+{
+	crypto_aead_ctx ctx;
+	crypto_aead_init_djb(&ctx, key, nonce);
+	crypto_aead_write(&ctx, cipher_text, mac, ad, ad_size,
+	                  plain_text, text_size);
+	crypto_wipe(&ctx, sizeof(ctx));
+}
+
+int crypto_aead_unlock_djb(u8 *plain_text, const u8  mac[16], const u8 key[32],
+                           const u8 nonce[8], const u8 *ad, size_t ad_size,
+                           const u8 *cipher_text, size_t text_size)
+{
+	crypto_aead_ctx ctx;
+	crypto_aead_init_djb(&ctx, key, nonce);
+	int mismatch = crypto_aead_read(&ctx, plain_text, mac, ad, ad_size,
+	                                cipher_text, text_size);
+	crypto_wipe(&ctx, sizeof(ctx));
+	return mismatch;
+}
+
 // vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/crypto.h b/src/crypto.h
index 44d4fe2..7f0d607 100644
--- a/src/crypto.h
+++ b/src/crypto.h
@@ -6,6 +6,20 @@
 #include "3p/monocypher/monocypher.h"
 #include "3p/monocypher/monocypher-rng.h"
 
+// -- SST-specific extensions to 4.0.1 API below --
+void crypto_aead_lock_djb(uint8_t       *cipher_text,
+                          uint8_t        mac  [16],
+                          const uint8_t  key  [32],
+                          const uint8_t  nonce[8],
+                          const uint8_t *ad,         size_t ad_size,
+                          const uint8_t *plain_text, size_t text_size);
+int crypto_aead_unlock_djb(uint8_t       *plain_text,
+                           const uint8_t  mac  [16],
+                           const uint8_t  key  [32],
+                           const uint8_t  nonce[8],
+                           const uint8_t *ad,          size_t ad_size,
+                           const uint8_t *cipher_text, size_t text_size);
+
 #endif
 
 // vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/src/democustom.c b/src/democustom.c
index 0ecbaa3..5dcbe01 100644
--- a/src/democustom.c
+++ b/src/democustom.c
@@ -14,9 +14,10 @@
  * PERFORMANCE OF THIS SOFTWARE.
  */
 
+#include <string.h>
+
 #include "bitbuf.h"
 #include "con_.h"
-#include "democustom.h"
 #include "demorec.h"
 #include "engineapi.h"
 #include "errmsg.h"
@@ -26,6 +27,8 @@
 #include "mem.h"
 #include "ppmagic.h"
 #include "vcall.h"
+#include "x86.h"
+#include "x86util.h"
 
 FEATURE()
 REQUIRE(demorec)
@@ -34,72 +37,64 @@ REQUIRE_GAMEDATA(vtidx_RecordPacket)
 
 static int nbits_msgtype, nbits_datalen;
 
-// The engine allows usermessages up to 255 bytes, we add 2 bytes of overhead,
-// and then there's the leading bits before that too (see create_message)
-static char bb_buf[DEMOCUSTOM_MSG_MAX + 4];
+// engine limit is 255, we use 2 bytes for header + round the bitstream to the
+// next whole byte, which gives 3 bytes overhead hence 252 here.
+#define CHUNKSZ 252
+
+static union {
+	char x[CHUNKSZ + /*7*/ 8]; // needs to be multiple of of 4!
+	bitbuf_cell _align; // just in case...
+} bb_buf;
 static struct bitbuf bb = {
-	bb_buf, sizeof(bb_buf), sizeof(bb_buf) * 8, 0, false, false, "SST"
+	{bb_buf.x}, sizeof(bb_buf), sizeof(bb_buf) * 8, 0, false, false, "SST"
 };
 
-static void create_message(struct bitbuf *msg, const void *buf, int len) {
-	// The way we pack our custom demo data is via a user message packet with
-	// type "HudText" - this causes the client to do a text lookup which will
-	// simply silently fail on invalid keys. By making the first byte null
-	// (creating an empty string), we get the rest of the packet to stick in
-	// whatever other data we want.
+static const void *createhdr(struct bitbuf *msg, int len, bool last) {
+	// We pack custom data into user message packets of type "HudText," with a
+	// leading null byte which the engine treats as an empty string. On demo
+	// playback, the client does a text lookup which fails silently on invalid
+	// keys, giving us the rest of the packet to stick in whatever data we want.
 	//
-	// Notes from Uncrafted:
-	// > But yeah the data you want to append is as follows:
-	// > - 6 bits (5 bits in older versions) for the message type - should be 23
-	// >   for user message
-	bitbuf_appendbits(msg, 23, nbits_msgtype);
-	// > - 1 byte for the user message type - should be 2 for HudText
-	bitbuf_appendbyte(msg, 2);
-	// > - ~~an int~~ 11 or 12 bits for the length of your data in bits,
-	bitbuf_appendbits(msg, len * 8, nbits_datalen); // NOTE: assuming len <= 254
-	// > - your data
-	// [first the aforementioned null byte, plus an arbitrary marker byte to
-	// avoid confusion when parsing the demo later...
-	bitbuf_appendbyte(msg, 0);
-	bitbuf_appendbyte(msg, 0xAC);
-	// ... and then just the data itself]
-	bitbuf_appendbuf(msg, buf, len);
-	// Thanks Uncrafted, very cool!
+	// Big thanks to our resident demo expert, Uncrafted, for explaining what to
+	// do here way back when this was first being figured out!
+	bitbuf_appendbits(msg, 23, nbits_msgtype); // type: 23 is user message
+	bitbuf_appendbyte(msg, 2); // user message type: 2 is HudText
+	bitbuf_appendbits(msg, len * 8, nbits_datalen); // our data length in bits
+	bitbuf_appendbyte(msg, 0); // aforementionied null byte
+	bitbuf_appendbyte(msg, 0xAC + last); // arbitrary marker byte to aid parsing
+	// store the data itself byte-aligned so there's no need to bitshift the
+	// universe (which would be both slower and more annoying to do)
+	bitbuf_roundup(msg);
+	return msg->buf + (msg->nbits >> 3);
 }
 
 typedef void (*VCALLCONV WriteMessages_func)(void *this, struct bitbuf *msg);
 static WriteMessages_func WriteMessages = 0;
 
 void democustom_write(const void *buf, int len) {
-	create_message(&bb, buf, len);
+	for (; len > CHUNKSZ; len -= CHUNKSZ) {
+		createhdr(&bb, CHUNKSZ, false);
+		memcpy(bb.buf + (bb.nbits >> 3), buf, CHUNKSZ);
+		bb.nbits += CHUNKSZ << 3;
+		WriteMessages(demorecorder, &bb);
+		bitbuf_reset(&bb);
+	}
+	createhdr(&bb, len, true);
+	memcpy(bb.buf + (bb.nbits >> 3), buf, len);
+	bb.nbits += len << 3;
 	WriteMessages(demorecorder, &bb);
 	bitbuf_reset(&bb);
 }
 
 static bool find_WriteMessages(void) {
-	// TODO(compat): rewrite this to just scan for a call instruction!
 	const uchar *insns = (*(uchar ***)demorecorder)[vtidx_RecordPacket];
-	// RecordPacket calls WriteMessages pretty much right away:
-	// 56           push  esi
-	// 57           push  edi
-	// 8B F1        mov   esi,ecx
-	// 8D BE        lea   edi,[esi + 0x68c]
-	// 8C 06 00 00
-	// 57           push  edi
-	// E8           call  CDemoRecorder_WriteMessages
-	// B0 EF FF FF
-	// So we just double check the byte pattern...
-	static const uchar bytes[] =
-#ifdef _WIN32
-		HEXBYTES(56, 57, 8B, F1, 8D, BE, 8C, 06, 00, 00, 57, E8);
-#else
-#warning This is possibly different on Linux too, have a look!
-		{-1, -1, -1, -1, -1, -1};
-#endif
-	if (!memcmp(insns, bytes, sizeof(bytes))) {
-		ssize off = mem_loadoffset(insns + sizeof(bytes));
-		WriteMessages = (WriteMessages_func)(insns + sizeof(bytes) + 4 + off);
-		return true;
+	// RecordPacket calls WriteMessages right away, so just look for a call
+	for (const uchar *p = insns; p - insns < 32;) {
+		if (*p == X86_CALL) {
+			WriteMessages = (WriteMessages_func)(p + 5 + mem_loadoffset(p + 1));
+			return true;
+		}
+		NEXT_INSN(p, "WriteMessages function");
 	}
 	return false;
 }
@@ -108,23 +103,21 @@ DECL_VFUNC_DYN(int, GetEngineBuildNumber)
 
 INIT {
 	// More UncraftedkNowledge:
-	// > yeah okay so [the usermessage length is] 11 bits if the demo protocol
-	// > is 11 or if the game is l4d2 and the network protocol is 2042.
-	// > otherwise it's 12 bits
-	// > there might be some other l4d2 versions where it's 11 but idk
+	// - usermessage length is:
+	//   - 11 bits in protocol 11, or l4d2 protocol 2042
+	//   - otherwise 12 bits
 	// So here we have to figure out the network protocol version!
 	// NOTE: assuming engclient != null as GEBN index relies on client version
 	int buildnum = GetEngineBuildNumber(engclient);
-	// condition is redundant until other GetEngineBuildNumber offsets are added
-	// if (GAMETYPE_MATCHES(L4D2)) {
+	//if (GAMETYPE_MATCHES(L4D2)) { // redundant until we add more GEBN offsets!
 		nbits_msgtype = 6;
 		// based on Some Code I Read, buildnum *should* be the protocol version,
 		// however L4D2 returns the actual game version instead, because sure
 		// why not. The only practical difference though is that the network
-		// protocol froze after 2042, so we just have to do a >=. No big deal
-		// really.
+		// protocol froze after 2042, so we just have to do a >=. Fair enough!
+		// TODO(compat): how does TLS affect this? no idea yet
 		if (buildnum >= 2042) nbits_datalen = 11; else nbits_datalen = 12;
-	// }
+	//}
 
 	return find_WriteMessages();
 }
diff --git a/src/democustom.h b/src/democustom.h
index a0a28b8..7f0c25e 100644
--- a/src/democustom.h
+++ b/src/democustom.h
@@ -17,12 +17,9 @@
 #ifndef INC_DEMOCUSTOM_H
 #define INC_DEMOCUSTOM_H
 
-/* maximum length of a custom demo message, in bytes */
-#define DEMOCUSTOM_MSG_MAX 253
-
 /*
- * Write a block of up to DEMOWRITER_MSG_MAX bytes into the currently recording
- * demo - NOT bounds checked, caller MUST ensure length is okay!
+ * Writes a custom demo message, automatically splitting into multiple demo
+ * packets if too long. Assumes a demo is currently being recorded.
  */
 void democustom_write(const void *buf, int len);
 
diff --git a/src/demorec.c b/src/demorec.c
index e176ab3..e728ca5 100644
--- a/src/demorec.c
+++ b/src/demorec.c
@@ -50,6 +50,10 @@ bool demorec_forceauto = false;
 #define SIGNONSTATE_SPAWN 5
 #define SIGNONSTATE_FULL 6
 
+DEF_PREDICATE(DemoControlAllowed, void)
+DEF_EVENT(DemoRecordStarting, void)
+DEF_EVENT(DemoRecordStopped, int)
+
 typedef void (*VCALLCONV SetSignonState_func)(void *, int);
 static SetSignonState_func orig_SetSignonState;
 static void VCALLCONV hook_SetSignonState(void *this_, int state) {
@@ -83,6 +87,9 @@ static void VCALLCONV hook_StopRecording(void *this) {
 		*recording = true;
 		*demonum = lastnum;
 	}
+	else {
+		EMIT_DemoRecordStopped(lastnum);
+	}
 }
 
 DECL_VFUNC_DYN(void, StartRecording)
@@ -90,10 +97,8 @@ DECL_VFUNC_DYN(void, StartRecording)
 static struct con_cmd *cmd_record, *cmd_stop;
 static con_cmdcb orig_record_cb, orig_stop_cb;
 
-DEF_PREDICATE(AllowDemoControl, void)
-
 static void hook_record_cb(const struct con_cmdargs *args) {
-	if (!CHECK_AllowDemoControl()) return;
+	if (!CHECK_DemoControlAllowed()) return;
 	bool was = *recording;
 	if (!was && args->argc == 2 || args->argc == 3) {
 		// safety check: make sure a directory exists, otherwise recording
@@ -152,10 +157,11 @@ static void hook_record_cb(const struct con_cmdargs *args) {
 		// mike: I think this is questionably necessary but I'm outvoted :)
 		con_msg("Demo recording started\n");
 	}
+	EMIT_DemoRecordStarting();
 }
 
 static void hook_stop_cb(const struct con_cmdargs *args) {
-	if (!CHECK_AllowDemoControl()) return;
+	if (!CHECK_DemoControlAllowed()) return;
 	wantstop = true;
 	orig_stop_cb(args);
 	wantstop = false;
@@ -229,6 +235,7 @@ bool demorec_start(const char *name) {
 	struct con_cmdargs args = {.argc = 2, .argv = {0, name, 0}};
 	orig_record_cb(&args);
 	if (!was && *recording) *demonum = 0; // same logic as in the hook
+	EMIT_DemoRecordStarting();
 	return *recording;
 }
 
@@ -237,11 +244,12 @@ int demorec_stop(void) {
 	// making this correct when recording and stopping in the menu lol
 	int ret = *demonum;
 	orig_StopRecording(demorecorder);
+	EMIT_DemoRecordStopped(ret);
 	return ret;
 }
 
-bool demorec_recording(void) {
-	return *recording;
+int demorec_demonum(void) {
+	return *recording ? *demonum : -1;
 }
 
 INIT {
diff --git a/src/demorec.h b/src/demorec.h
index 2de4f24..b53896f 100644
--- a/src/demorec.h
+++ b/src/demorec.h
@@ -52,21 +52,38 @@ bool demorec_start(const char *name);
 /*
  * Stops recording the current demo and returns the number of demos recorded
  * (the first will have the original basename + .dem extension; the rest will
- * have the _N.dem suffixes).
+ * have the _N.dem suffixes). Value will be zero if the recording is stopped
+ * before the game has even gotten a chance to create the first demo file.
  */
 int demorec_stop(void);
 
 /*
- * Queries whether a demo is currently being recorded.
+ * Returns the current number in the recording sequence, or -1 if not recording.
+ * Value may be 0 if recording was requested but has yet to start (say, because
+ * we have yet to join a map).
  */
-bool demorec_recording(void);
+int demorec_demonum(void);
 
 /*
  * Used to determine whether to allow usage of the normal record and stop
  * commands. Code which takes over control of demo recording can use this to
  * block the user from interfering.
  */
-DECL_PREDICATE(AllowDemoControl)
+DECL_PREDICATE(DemoControlAllowed, void)
+
+/*
+ * Emitted whenever a recording session is about to be started, as a result of
+ * either the record command or a call to the demorec_start() function. A demo
+ * file won't actually have been created yet; this merely indicates that a
+ * request to record has happened.
+ */
+DECL_EVENT(DemoRecordStarting, void)
+
+/*
+ * Emitted when the current demo or series of demos has finished recording.
+ * Receives the number of recorded demo files (which could be 0) as an argument.
+ */
+DECL_EVENT(DemoRecordStopped, int)
 
 #endif
 
diff --git a/src/engineapi.c b/src/engineapi.c
index f4a54d6..24a2d6b 100644
--- a/src/engineapi.c
+++ b/src/engineapi.c
@@ -41,6 +41,7 @@ DECL_VFUNC(void *, GetGlobalVars, 1) // seems to be very stable, thank goodness
 void *globalvars;
 
 void *inputsystem, *vgui;
+struct CServerPlugin *pluginhandler;
 
 DECL_VFUNC_DYN(void *, GetAllServerClasses)
 
@@ -48,6 +49,7 @@ DECL_VFUNC_DYN(void *, GetAllServerClasses)
 
 bool engineapi_init(int pluginver) {
 	if (!con_detect(pluginver)) return false;
+	pluginhandler = factory_engine("ISERVERPLUGINHELPERS001", 0);
 
 	if (engclient = factory_engine("VEngineClient015", 0)) {
 		_gametype_tag |= _gametype_tag_Client015;
diff --git a/src/engineapi.h b/src/engineapi.h
index d740c8c..fbc062b 100644
--- a/src/engineapi.h
+++ b/src/engineapi.h
@@ -136,6 +136,23 @@ extern struct VEngineServer *engserver;
 extern void *globalvars;
 extern void *inputsystem, *vgui;
 
+// XXX: not exactly engine *API* but not curently clear where else to put this
+struct CPlugin {
+	char description[128];
+	bool paused;
+	void *theplugin; // our own "this" pointer (or whichever other plugin it is)
+	int ifacever;
+	// should be the plugin library, but in old Source branches it's just null,
+	// because CServerPlugin::Load() erroneously shadows this field with a local
+	void *module;
+};
+struct CServerPlugin /* : IServerPluginHelpers */ {
+	void **vtable;
+	struct CUtlVector plugins;
+	/*IPluginHelpersCheck*/ void *pluginhlpchk;
+};
+extern struct CServerPlugin *pluginhandler;
+
 /*
  * Called on plugin init to attempt to initialise various core interfaces.
  * This includes console/cvar initialisation and populating gametype and
diff --git a/src/os-win32.h b/src/os-win32.h
index a006083..db20964 100644
--- a/src/os-win32.h
+++ b/src/os-win32.h
@@ -50,10 +50,10 @@ typedef unsigned short os_char;
 
 #define OS_MAIN wmain
 
-static inline void *os_dlopen(const ushort *name) {
+static inline void *os_dlopen(const unsigned short *name) {
 	return LoadLibraryW(name);
 }
-static inline void *os_dlhandle(const ushort *name) {
+static inline void *os_dlhandle(const unsigned short *name) {
 	return GetModuleHandleW(name);
 }
 static inline void *os_dlsym(void *m, const char *s) {
diff --git a/src/rinput.c b/src/rinput.c
index 6b6d4d7..79b6661 100644
--- a/src/rinput.c
+++ b/src/rinput.c
@@ -64,7 +64,7 @@ DEF_CVAR_UNREG(m_rawinput, "Use Raw Input for mouse input (SST reimplementation)
 DEF_CVAR_MINMAX(sst_mouse_factor, "Number of hardware mouse counts per step",
 		1, 1, 20, /*CON_ARCHIVE |*/ CON_HIDDEN)
 
-static ssize __stdcall inproc(void *wnd, uint msg, ssize wp, ssize lp) {
+static ssize __stdcall inproc(void *wnd, uint msg, usize wp, ssize lp) {
 	switch (msg) {
 		case WM_INPUT:;
 			char buf[sizeof(RAWINPUTHEADER) + sizeof(RAWMOUSE) /* = 40 */];
diff --git a/src/sst.c b/src/sst.c
index 61a021f..9fd1108 100644
--- a/src/sst.c
+++ b/src/sst.c
@@ -30,6 +30,7 @@
 #include "gametype.h"
 #include "hook.h"
 #include "os.h"
+#include "sst.h"
 #include "vcall.h"
 #include "version.h"
 
@@ -46,10 +47,13 @@ static int ifacever;
 void *clientlib = 0;
 
 bool sst_earlyloaded = false; // see deferinit() below
+bool sst_userunloaded = false; // see hook_plugin_unload_cb() below
+
+#define VDFBASENAME "SourceSpeedrunTools"
 
 #ifdef _WIN32
 extern long __ImageBase; // this is actually the PE header struct but don't care
-#define ownhandle() ((void *)&__ImageBase)
+static inline void *ownhandle(void) { return &__ImageBase; }
 #else
 // sigh, _GNU_SOURCE crap. define here instead >:(
 typedef struct {
@@ -59,16 +63,16 @@ typedef struct {
 	void *dli_saddr;
 } Dl_info;
 int dladdr1(const void *addr, Dl_info *info, void **extra_info, int flags);
-static inline void *ownhandle(void) {
+static void *ownhandle(void) {
+	static void *cached = 0;
 	Dl_info dontcare;
-	void *dl;
-	dladdr1((void *)&ownhandle, &dontcare, &dl, /*RTLD_DL_LINKMAP*/ 2);
-	return dl;
+	if (!cached) {
+		dladdr1((void *)&ownhandle, &dontcare, &cached, /*RTLD_DL_LINKMAP*/ 2);
+	}
+	return cached;
 }
 #endif
 
-#define VDFBASENAME "SourceSpeedrunTools"
-
 #ifdef _WIN32
 // not a proper check, just a short-circuit check to avoid doing more work.
 static inline bool checksamedrive(const ushort *restrict path1,
@@ -305,6 +309,45 @@ e:	con_warn("!!! SOME FEATURES MAY BE BROKEN !!!\n");
 	return false;
 }
 
+DEF_PREDICATE(AllowPluginLoading, bool)
+DEF_EVENT(PluginLoaded, void)
+DEF_EVENT(PluginUnloaded, void)
+
+static struct con_cmd *cmd_plugin_load, *cmd_plugin_unload;
+static con_cmdcb orig_plugin_load_cb, orig_plugin_unload_cb;
+
+static int ownidx; // XXX: super hacky way of getting this to do_unload()
+
+static void hook_plugin_load_cb(const struct con_cmdargs *args) {
+	if (args->argc == 1) return;
+	if (!CHECK_AllowPluginLoading(true)) return;
+	orig_plugin_load_cb(args);
+	EMIT_PluginLoaded();
+}
+static void hook_plugin_unload_cb(const struct con_cmdargs *args) {
+	if (args->argc == 1) return;
+	if (!CHECK_AllowPluginLoading(false)) return;
+	int idx = atoi(args->argv[1]);
+	struct CPlugin **plugins = pluginhandler->plugins.m.mem;
+	if (idx >= 0 && idx < pluginhandler->plugins.sz &&
+			plugins[idx]->theplugin == &plugin_obj) {
+		sst_userunloaded = true;
+		ownidx = idx;
+#ifdef __clang__
+		// thanks clang for forcing use of return here and THEN warning about it
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wpedantic"
+		__attribute__((musttail)) return orig_plugin_unload_cb(args);
+#pragma clang diagnostic pop
+#else
+#error We are tied to clang without an assembly solution for this!
+#endif
+	}
+	// if it's some other plugin being unloaded, we can keep doing stuff after
+	orig_plugin_unload_cb(args);
+	EMIT_PluginUnloaded();
+}
+
 static bool do_load(ifacefactory enginef, ifacefactory serverf) {
 	if (!hook_init()) {
 		errmsg_warnsys("couldn't set up memory for function hooking");
@@ -326,46 +369,34 @@ static bool do_load(ifacefactory enginef, ifacefactory serverf) {
 	*p++ = (void *)&nop_p_v;		  // OnEdictAllocated
 	*p   = (void *)&nop_p_v;		  // OnEdictFreed
 	if (!deferinit()) { do_featureinit(); fixes_apply(); }
+	if (pluginhandler) {
+		cmd_plugin_load = con_findcmd("plugin_load");
+		orig_plugin_load_cb = cmd_plugin_load->cb;
+		cmd_plugin_load->cb = &hook_plugin_load_cb;
+		cmd_plugin_unload = con_findcmd("plugin_unload");
+		orig_plugin_unload_cb = cmd_plugin_unload->cb;
+		cmd_plugin_unload->cb = &hook_plugin_unload_cb;
+	}
 	return true;
 }
 
-struct CServerPlugin /* : IServerPluginHelpers */ {
-	void **vtable;
-	struct CUtlVector plugins;
-	/*IPluginHelpersCheck*/ void *pluginhlpchk;
-};
-struct CPlugin {
-	char description[128];
-	bool paused;
-	void *theplugin; // our own "this" pointer (or whichever other plugin it is)
-	int ifacever;
-	// should be the plugin library, but in old Source branches it's just null,
-	// because CServerPlugin::Load() erroneously shadows this field with a local
-	void *module;
-};
-
 static void do_unload(void) {
 #ifdef _WIN32 // this is only relevant in builds that predate linux support
-	struct CServerPlugin *pluginhandler =
-			factory_engine("ISERVERPLUGINHELPERS001", 0);
-	if (pluginhandler) { // if not, oh well too bad we tried :^)
-		struct CPlugin **plugins = pluginhandler->plugins.m.mem;
-		int n = pluginhandler->plugins.sz;
-		for (struct CPlugin **pp = plugins; pp - plugins < n; ++pp) {
-			if ((*pp)->theplugin == (void *)&plugin_obj) {
-				// see comment in CPlugin above. setting this to the real handle
-				// right before the engine tries to unload us allows it to
-				// actually do so. in newer branches this is redundant but
-				// doesn't do any harm so it's just unconditional.
-				// NOTE: old engines ALSO just leak the handle and never call
-				// Unload() if Load() fails; can't really do anything about that
-				(*pp)->module = ownhandle();
-				break;
-			}
+	if (pluginhandler) { // if not, oh well too bad :^)
+		cmd_plugin_load->cb = orig_plugin_load_cb;
+		cmd_plugin_unload->cb = orig_plugin_unload_cb;
+		if (sst_userunloaded) {
+			struct CPlugin **plugins = pluginhandler->plugins.m.mem;
+			// see comment in CPlugin above. setting this to the real handle
+			// right before the engine tries to unload us allows it to actually
+			// do so. in newer branches this is redundant but doesn't do any
+			// harm so it's just unconditional. NOTE: old engines ALSO just leak
+			// the handle and never call Unload() if Load() fails; can't really
+			// do anything about that.
+			plugins[ownidx]->module = ownhandle();
 		}
 	}
 #endif
-
 	endfeatures();
 #ifdef __linux__
 	if (clientlib) dlclose(clientlib);
diff --git a/src/sst.h b/src/sst.h
index 8c10798..f7e4bb4 100644
--- a/src/sst.h
+++ b/src/sst.h
@@ -24,10 +24,17 @@
 DECL_EVENT(ClientActive, struct edict */*player*/)
 DECL_EVENT(Tick, bool /*simulating*/)
 
+DECL_PREDICATE(AllowPluginLoading, void)
+DECL_EVENT(PluginLoaded, void)
+DECL_EVENT(PluginUnloaded, void)
+
 extern void *clientlib;
 
 /* occasionally useful: quick query to determine how sst was loaded */
 extern bool sst_earlyloaded;
+/* similar query for how we are being unloaded - ONLY valid during unload */
+// TODO(opt): we can skip a whole bunch of cleanup when exiting the game!
+extern bool sst_userunloaded;
 
 #endif
 
diff --git a/test/bitbuf.test.c b/test/bitbuf.test.c
index 324a7f6..554d3a2 100644
--- a/test/bitbuf.test.c
+++ b/test/bitbuf.test.c
@@ -15,15 +15,11 @@ static union {
 static struct bitbuf bb = {bb_buf.buf, 512, 512 * 8, 0, false, false, "test"};
 
 TEST("The possible UB in bitbuf_appendbuf shouldn't trigger horrible bugs") {
-	char unalign[3] = {'X', 'X', 'X'};
-	char _buf[32 + sizeof(bitbuf_cell)];
-	char *buf = _buf;
-	if (bitbuf_align <= 1) {
-		// *shouldn't* happen
+	if (bitbuf_align <= 1) { // *shouldn't* happen
 		fputs("what's going on with the alignment???\n", stderr);
 		return false;
 	}
-	// make sure the pointer is definitely misaligned
+	char _buf[32 + _Alignof(bitbuf_cell)], *buf = _buf;
 	while (!((usize)buf % bitbuf_align)) ++buf;
 
 	memcpy(buf, "Misaligned test buffer contents!", 32);
@@ -31,4 +27,18 @@ TEST("The possible UB in bitbuf_appendbuf shouldn't trigger horrible bugs") {
 	return !memcmp(bb.buf, buf, 32);
 }
 
+TEST("Aligning to the next byte should work as intended") {
+	for (int i = 0; i < 65535; i += 8) {
+		bb.curbit = i;
+		bitbuf_roundup(&bb);
+		if (bb.curbit != i) return false; // don't round if already rounded
+		for (int j = i + 1; j < i + 8; ++j) {
+			bb.curbit = j;
+			bitbuf_roundup(&bb);
+			if (bb.curbit != i + 8) return false;
+		}
+	}
+	return true;
+}
+
 // vi: sw=4 ts=4 noet tw=80 cc=80
diff --git a/tools/genkeypair.c b/tools/genkeypair.c
new file mode 100644
index 0000000..7db856b
--- /dev/null
+++ b/tools/genkeypair.c
@@ -0,0 +1,42 @@
+// *SUPER* low effort x25519 keygen tool, for lack of better tool on hand.
+// To compile:
+// Unix: $CC -O2 -Dbool=_Bool -o.build/genkeypair tools/genkeypair.c
+// Windows: clang-cl -fuse-ld=lld -O2 -Dbool=_Bool -Fe.build/genkeypair.exe tools/genkeypair.c /link advapi32.lib
+
+#include <stdio.h>
+#include <stdlib.h>
+
+#include "../src/3p/monocypher/monocypher.c" // yeah, we're doing this.
+#include "../src/os.h"
+
+int main(void) {
+	unsigned char prv[32], pub[32];
+	os_randombytes(prv, sizeof(prv));
+	crypto_x25519_public_key(pub, prv);
+	// and now the worst string formatting code you've ever seen!!!
+	fputs("Private key: ", stdout);
+	for (int i = 0; i < sizeof(prv); i += 4) {
+		fprintf(stdout, "%.2X%.2X%.2X%.2X",
+				prv[i], prv[i + 1], prv[i + 2], prv[i + 3]);
+	}
+	fputs("\n{", stdout);
+	for (int i = 0; i < sizeof(prv) - 4; i += 4) {
+		fprintf(stdout, "0x%.2X, 0x%.2X, 0x%.2X, 0x%.2X, ",
+				prv[i], prv[i + 1], prv[i + 2], prv[i + 3]);
+	}
+	fprintf(stdout, "0x%.2X, 0x%.2X, 0x%.2X, 0x%.2X}\n\nPublic key:  ",
+			prv[sizeof(prv) - 4], prv[sizeof(prv) - 3],
+			prv[sizeof(prv) - 2], prv[sizeof(prv) - 1]);
+	for (int i = 0; i < sizeof(pub); i += 4) {
+		fprintf(stdout, "%.2X%.2X%.2X%.2X",
+				pub[i], pub[i + 1], pub[i + 2], pub[i + 3]);
+	}
+	fputs("\n{", stdout);
+	for (int i = 0; i < sizeof(pub) - 4; i += 4) {
+		fprintf(stdout, "0x%.2X, 0x%.2X, 0x%.2X, 0x%.2X, ",
+				pub[i], pub[i + 1], pub[i + 2], pub[i + 3]);
+	}
+	fprintf(stdout, "0x%.2X, 0x%.2X, 0x%.2X, 0x%.2X}\n",
+			pub[sizeof(pub) - 4], pub[sizeof(pub) - 3],
+			pub[sizeof(pub) - 2], pub[sizeof(pub) - 1]);
+}