// coop_audio_validate -- quantify the fidelity of the injection audio-capture path.
//
// "The audio sounds slightly off" is hard to act on; this turns it into numbers. It
// plays a known sine tone (coop_tone), injects coop_hook.dll exactly as the host does
// (late attach: the ring is created after injection, so the hook must guess+measure the
// rate -- the Brotato/Godot case), captures the hook's ring into memory, and runs the
// fidelity analyzer (coop/tone_analysis.hpp): pitch error in cents, SNR/THD, click +
// dropout counts. It also writes the captured audio to a .wav so it can be *listened* to.
//
//   coop_audio_validate                 # spawn coop_tone @ 44100 Hz / 1000 Hz, full self-test
//   coop_audio_validate --rate 48000 --freq 440 --seconds 8
//   coop_audio_validate --pid 1234 --freq 1000        # attach to an already-running tone/game
//   coop_audio_validate --wav capture.wav --freq 1000 # just analyze a recorded .wav (e.g. a
//                                                      # host render-output dump from real Brotato)
//
// Run from bin/<config>/tools/ (next to the deployable root that holds coop_hook.dll;
// coop_tone.exe is found in the sibling tests/ folder). The hook trace is %TEMP%\coop_hook.log.

#include <algorithm>
#include <atomic>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <string>
#include <thread>
#include <vector>

#include <mutex>

#include <windows.h>

#include <mmreg.h>
#include <objbase.h>

#include "audio/audio_loopback.hpp"
#include "audio/process_loopback_capture.hpp"
#include "coop/audio_ring.hpp"
#include "coop/protocol.hpp"
#include "coop/shared_memory.hpp"
#include "coop/tone_analysis.hpp"
#include "coop/tool_paths.hpp"
#include "coop/wav.hpp"
#include "tone_source.hpp" // in-process sine renderer (shared with coop_tone), for --selfcheck

namespace {

struct Options {
	unsigned long pid = 0;    // attach to this pid instead of spawning coop_tone
	unsigned long listen = 0; // --listen: passively loopback-capture this pid's output (e.g. coop_host)
	double freq = 1000.0;     // the tone frequency (for pitch analysis)
	unsigned rate = 44100;    // tone render rate (the Brotato/Godot non-device case by default)
	unsigned channels = 2;
	unsigned bits = 32;     // 32 = float, 16 = pcm
	int seconds = 6;        // capture duration
	bool render = false;    // --render: measure the host RENDER path (run_hooked), not just capture
	bool baseline = false;  // --baseline: loopback-capture the tone directly (no hook/mirror) as a floor
	bool selfcheck = false; // --selfcheck: render a clean tone in-process + self-capture (control for self-capture)
	std::wstring wav_in;    // analyze this .wav instead of capturing
	std::wstring wav_out;   // where to dump the captured audio (default next to the exe)
};

std::wstring sibling(const std::wstring& path, const wchar_t* name)
{
	const std::size_t slash = path.find_last_of(L"\\/");
	return (slash == std::wstring::npos ? std::wstring() : path.substr(0, slash + 1)) + name;
}

// coop_tone.exe is staged in bin/<config>/tests/; this tool runs from bin/<config>/tools/.
std::wstring find_coop_tone()
{
	const std::wstring here = coop::exe_directory() + L"coop_tone.exe";
	if (GetFileAttributesW(here.c_str()) != INVALID_FILE_ATTRIBUTES) {
		return here;
	}
	std::wstring dir = coop::exe_directory();
	if (!dir.empty()) {
		dir.pop_back();
	}
	const std::size_t slash = dir.find_last_of(L"\\/");
	const std::wstring root = (slash == std::wstring::npos) ? std::wstring() : dir.substr(0, slash + 1);
	const std::wstring in_tests = root + L"tests\\coop_tone.exe";
	if (GetFileAttributesW(in_tests.c_str()) != INVALID_FILE_ATTRIBUTES) {
		return in_tests;
	}
	return here;
}

// --- injection (mirrors coop_audio_probe, incl. the x86 WOW64 helper) -------------------

bool inject_via_helper(unsigned long pid, const std::wstring& dll_path)
{
	const std::wstring helper = sibling(dll_path, L"coop_inject_x86.exe");
	const std::wstring x86_dll = sibling(dll_path, L"coop_hook_x86.dll");
	if (GetFileAttributesW(helper.c_str()) == INVALID_FILE_ATTRIBUTES
	    || GetFileAttributesW(x86_dll.c_str()) == INVALID_FILE_ATTRIBUTES) {
		std::printf("ERROR: x86 helper/dll missing next to the tool.\n");
		return false;
	}
	std::wstring cmd = L"\"" + helper + L"\" " + std::to_wstring(pid) + L" \"" + x86_dll + L"\"";
	STARTUPINFOW si{};
	si.cb = sizeof(si);
	PROCESS_INFORMATION pi{};
	if (!CreateProcessW(helper.c_str(), cmd.data(), nullptr, nullptr, FALSE, 0, nullptr, nullptr, &si, &pi)) {
		std::printf("ERROR: CreateProcess(coop_inject_x86) failed (%lu).\n", GetLastError());
		return false;
	}
	WaitForSingleObject(pi.hProcess, INFINITE);
	DWORD code = 1;
	GetExitCodeProcess(pi.hProcess, &code);
	CloseHandle(pi.hThread);
	CloseHandle(pi.hProcess);
	return code == 0;
}

bool inject(unsigned long pid, const std::wstring& dll_path)
{
	if (GetFileAttributesW(dll_path.c_str()) == INVALID_FILE_ATTRIBUTES) {
		std::printf("ERROR: coop_hook.dll not found next to the tool.\n");
		return false;
	}
	const DWORD access =
	    PROCESS_CREATE_THREAD | PROCESS_QUERY_INFORMATION | PROCESS_VM_OPERATION | PROCESS_VM_WRITE | PROCESS_VM_READ;
	HANDLE process = OpenProcess(access, FALSE, pid);
	if (process == nullptr) {
		std::printf("ERROR: OpenProcess(%lu) failed (%lu). Run as administrator?\n", pid, GetLastError());
		return false;
	}
	USHORT proc_machine = IMAGE_FILE_MACHINE_UNKNOWN, native_machine = IMAGE_FILE_MACHINE_UNKNOWN;
	if (IsWow64Process2(process, &proc_machine, &native_machine) && proc_machine != IMAGE_FILE_MACHINE_UNKNOWN) {
		CloseHandle(process);
		return inject_via_helper(pid, dll_path);
	}
	const SIZE_T bytes = (dll_path.size() + 1) * sizeof(wchar_t);
	void* remote = VirtualAllocEx(process, nullptr, bytes, MEM_COMMIT | MEM_RESERVE, PAGE_READWRITE);
	bool ok = false;
	if (remote != nullptr && WriteProcessMemory(process, remote, dll_path.c_str(), bytes, nullptr)) {
		auto load_library =
		    reinterpret_cast<LPTHREAD_START_ROUTINE>(GetProcAddress(GetModuleHandleW(L"kernel32.dll"), "LoadLibraryW"));
		HANDLE thread = CreateRemoteThread(process, nullptr, 0, load_library, remote, 0, nullptr);
		if (thread != nullptr) {
			WaitForSingleObject(thread, INFINITE);
			DWORD exit_code = 0;
			GetExitCodeThread(thread, &exit_code);
			CloseHandle(thread);
			ok = (exit_code != 0);
		}
	}
	if (remote != nullptr) {
		VirtualFreeEx(process, remote, 0, MEM_RELEASE);
	}
	CloseHandle(process);
	return ok;
}

void enable_hook_trace()
{
	wchar_t dir[MAX_PATH] = {};
	if (GetTempPathW(MAX_PATH, dir) != 0) {
		const std::wstring sentinel = std::wstring(dir) + L"coop_hook.log.on";
		HANDLE h = CreateFileW(sentinel.c_str(), GENERIC_WRITE, FILE_SHARE_READ, nullptr, OPEN_ALWAYS,
		                       FILE_ATTRIBUTE_NORMAL, nullptr);
		if (h != INVALID_HANDLE_VALUE) {
			CloseHandle(h);
		}
	}
}

// Spawn coop_tone at the requested format; parse "TONE_RENDERING pid=NNN ..." from its
// stdout. Returns the tone process + its pid (0 on failure). We keep the handle so the
// tone keeps playing for the whole capture and is killed at the end.
HANDLE spawn_tone(const Options& o, unsigned long& tone_pid)
{
	const std::wstring exe = find_coop_tone();
	if (GetFileAttributesW(exe.c_str()) == INVALID_FILE_ATTRIBUTES) {
		std::printf("ERROR: coop_tone.exe not found (looked next to the tool and in ../tests/).\n");
		return nullptr;
	}
	HANDLE rd = nullptr, wr = nullptr;
	SECURITY_ATTRIBUTES sa{sizeof(sa), nullptr, TRUE};
	if (!CreatePipe(&rd, &wr, &sa, 0)) {
		return nullptr;
	}
	SetHandleInformation(rd, HANDLE_FLAG_INHERIT, 0);

	// coop_tone [seconds] [freq] [rate] [channels] [bits] [float|pcm]
	const wchar_t* kind = (o.bits == 32) ? L"float" : L"pcm";
	std::wstring cmd = L"\"" + exe + L"\" " + std::to_wstring(o.seconds + 4) + L" "
	                   + std::to_wstring(static_cast<long>(o.freq)) + L" " + std::to_wstring(o.rate) + L" "
	                   + std::to_wstring(o.channels) + L" " + std::to_wstring(o.bits) + L" " + kind;
	STARTUPINFOW si{};
	si.cb = sizeof(si);
	si.dwFlags = STARTF_USESTDHANDLES;
	si.hStdOutput = wr;
	si.hStdError = wr;
	PROCESS_INFORMATION pi{};
	const BOOL launched =
	    CreateProcessW(exe.c_str(), cmd.data(), nullptr, nullptr, TRUE, 0, nullptr, nullptr, &si, &pi);
	CloseHandle(wr);
	if (!launched) {
		std::printf("ERROR: CreateProcess(coop_tone) failed (%lu).\n", GetLastError());
		CloseHandle(rd);
		return nullptr;
	}
	CloseHandle(pi.hThread);
	tone_pid = pi.dwProcessId;

	// Read the first line ("TONE_RENDERING ...") so we know audio is actually flowing.
	std::string line;
	char ch = 0;
	DWORD got = 0;
	const DWORD start = GetTickCount();
	while (GetTickCount() - start < 5000) {
		if (ReadFile(rd, &ch, 1, &got, nullptr) && got == 1) {
			if (ch == '\n') {
				break;
			}
			if (ch != '\r') {
				line.push_back(ch);
			}
		} else {
			break;
		}
	}
	CloseHandle(rd);
	if (line.rfind("TONE_RENDERING", 0) == 0) {
		std::printf("coop_tone: %s\n", line.c_str());
		return pi.hProcess;
	}
	std::printf("ERROR: coop_tone did not start rendering (got: \"%s\").\n", line.c_str());
	TerminateProcess(pi.hProcess, 1);
	CloseHandle(pi.hProcess);
	tone_pid = 0;
	return nullptr;
}

// Capture the hook's audio ring for `seconds`, draining frequently so the tool itself
// never causes an overrun -- the captured buffer is then exactly what the hook produced.
// Fills `pcm` (interleaved) and reports the declared format. Returns false if no format.
bool capture_ring(coop::AudioRingHeader* ring, int seconds, std::vector<std::uint8_t>& pcm, std::uint32_t& rate,
                  std::uint32_t& channels, std::uint32_t& bits, std::uint32_t& format_tag, std::uint64_t& overruns)
{
	// Wait up to 8 s for the hook to publish a format (late attach measures the rate first).
	const DWORD wait_end = GetTickCount() + 8000;
	while (!coop::audio_ring_format_ready(*ring)) {
		if (GetTickCount() >= wait_end) {
			std::printf("ERROR: hook never published an audio format (no stream captured).\n");
			return false;
		}
		Sleep(20);
	}
	rate = ring->sample_rate;
	channels = ring->channels;
	bits = ring->bits;
	format_tag = ring->format_tag;
	std::printf("Hook published format: %u Hz / %u ch / %u-bit / tag %u. Capturing %d s...\n", rate, channels, bits,
	            format_tag, seconds);

	std::vector<std::uint8_t> scratch(coop::kAudioRingCapacity);
	const DWORD cap_end = GetTickCount() + static_cast<DWORD>(seconds) * 1000;
	while (GetTickCount() < cap_end) {
		std::uint32_t got = coop::audio_ring_pop(*ring, scratch.data(), static_cast<std::uint32_t>(scratch.size()));
		if (got > 0) {
			pcm.insert(pcm.end(), scratch.begin(), scratch.begin() + got);
		} else {
			Sleep(2); // ring momentarily empty; poll again shortly
		}
	}
	// Drain any tail.
	std::uint32_t got = 0;
	while ((got = coop::audio_ring_pop(*ring, scratch.data(), static_cast<std::uint32_t>(scratch.size()))) > 0) {
		pcm.insert(pcm.end(), scratch.begin(), scratch.begin() + got);
	}
	overruns = ring->overruns.load(std::memory_order_relaxed);
	return !pcm.empty();
}

void print_report(const coop::ToneReport& r, double expected_hz, std::uint32_t declared_rate, std::uint64_t overruns)
{
	std::printf("\n================ FIDELITY REPORT ================\n");
	std::printf("  samples analyzed : %zu frames (%.2f s @ %u Hz)\n", r.frames, r.duration_sec, r.sample_rate);
	std::printf("  level            : RMS %.4f  peak %.4f  clipped %.3f%%\n", r.rms, r.peak,
	            r.clipped_fraction * 100.0);
	if (expected_hz > 0.0) {
		std::printf("  PITCH            : %.2f Hz captured vs %.2f Hz played  ->  %+.1f cents (x%.4f)\n", r.dominant_hz,
		            expected_hz, r.pitch_error_cents, r.pitch_error_ratio);
		// If the pitch is off, the most likely cause is a wrong declared rate. Show the rate
		// the captured pitch implies, so a misdetection is obvious at a glance.
		if (r.pitch_error_ratio > 0.0) {
			const double implied_true_rate = declared_rate / r.pitch_error_ratio;
			std::printf("  implied true rate: ~%.0f Hz (declared %u Hz)%s\n", implied_true_rate, declared_rate,
			            std::fabs(r.pitch_error_cents) > 15.0 ? "   <-- MISMATCH" : "");
		}
		std::printf("  spectral purity  : SNR %.1f dB   THD %.3f%%\n", r.snr_db, r.thd_percent);
	}
	std::printf("  discontinuities  : %u clicks (%.2f/s)\n", r.glitch_count, r.glitch_rate_per_sec);
	std::printf("  dropouts         : %u gaps, %.1f ms total\n", r.dropout_count, r.dropout_ms);
	if (overruns != UINT64_MAX) {
		std::printf("  ring overruns    : %llu (host fell behind -> dropped packets)\n",
		            static_cast<unsigned long long>(overruns));
	}

	std::printf("------------------- VERDICT --------------------\n");
	int problems = 0;
	if (expected_hz > 0.0 && std::fabs(r.pitch_error_cents) > 15.0) {
		std::printf("  [X] PITCH SHIFT: captured rate is wrong (audible). Likely a mis-measured\n"
		            "      late-attach rate -- see implied true rate above.\n");
		++problems;
	}
	if (r.dropout_count > 0) {
		std::printf("  [X] DROPOUTS: %u silence gap(s) -- choppy / 'metallic' under-run artifacts.\n", r.dropout_count);
		++problems;
	}
	if (r.glitch_rate_per_sec > 1.0) {
		std::printf("  [X] CLICKS: %.1f discontinuities/s -- torn/dropped packets.\n", r.glitch_rate_per_sec);
		++problems;
	}
	if (expected_hz > 0.0 && r.snr_db < 40.0) {
		std::printf("  [X] DISTORTION: SNR %.1f dB is low for a pure tone.\n", r.snr_db);
		++problems;
	}
	if (problems == 0) {
		std::printf("  [OK] Captured audio is faithful (pitch, purity, continuity all good).\n");
	}
	std::printf("=================================================\n");
}

// Resolve a WAVEFORMATEX (possibly EXTENSIBLE) to the scalar fields the analyzer wants.
void resolve_waveformat(const WAVEFORMATEX* w, std::uint32_t& rate, std::uint32_t& channels, std::uint32_t& bits,
                        std::uint32_t& tag)
{
	rate = w->nSamplesPerSec;
	channels = w->nChannels;
	bits = w->wBitsPerSample;
	tag = w->wFormatTag;
	if (w->wFormatTag == WAVE_FORMAT_EXTENSIBLE && w->cbSize >= 22) {
		const auto* ext = reinterpret_cast<const WAVEFORMATEXTENSIBLE*>(w);
		tag = (ext->SubFormat == KSDATAFORMAT_SUBTYPE_IEEE_FLOAT) ? coop::kToneFormatFloat : coop::kToneFormatPcm;
	}
}

// Loopback-capture `pid`'s render output (device-clock faithful, gaps included) for
// `seconds`, into `pcm`, and report the device format. Shared by --render (self) and
// --baseline (the tone directly). Assumes COM is already initialized on this thread.
bool loopback_capture_pid(unsigned long pid, int seconds, std::vector<std::uint8_t>& pcm, std::uint32_t& rate,
                          std::uint32_t& channels, std::uint32_t& bits, std::uint32_t& tag, std::uint32_t& block_align)
{
	WAVEFORMATEX* mix = coop::default_render_format();
	if (mix == nullptr) {
		std::printf("ERROR: could not get the default render format.\n");
		return false;
	}
	resolve_waveformat(mix, rate, channels, bits, tag);
	block_align = mix->nBlockAlign;
	std::mutex m;
	coop::ProcessLoopbackCapture cap;
	const bool ok = cap.start(pid, mix, [&](const BYTE* data, std::uint32_t frames, bool silent) {
		const std::size_t bytes = static_cast<std::size_t>(frames) * mix->nBlockAlign;
		std::lock_guard<std::mutex> lk(m);
		if (silent || data == nullptr) {
			pcm.insert(pcm.end(), bytes, 0);
		} else {
			pcm.insert(pcm.end(), data, data + bytes);
		}
	});
	if (ok) {
		Sleep(static_cast<DWORD>(seconds) * 1000);
	}
	cap.stop();
	CoTaskMemFree(mix);
	return ok;
}

// --listen: passively loopback-capture an already-running process's render output (e.g. the
// live coop_host while it mirrors a real game). On the hooked path the game is silenced, so
// coop_host's render mix IS exactly what the guest hears -- this records it to a .wav and runs
// the analyzer. Pass --freq for an in-game test tone to get pitch numbers; otherwise the level /
// click / dropout metrics still apply to real game audio. No injection, no mirror -- just listen.
int run_listen_mode(const Options& o)
{
	const bool com_ok = SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED));
	std::printf("Listening to pid %lu's render output for %d s (e.g. the live coop_host mirror)...\n", o.listen,
	            o.seconds);
	std::vector<std::uint8_t> pcm;
	std::uint32_t rate = 0, channels = 0, bits = 0, tag = 0, block = 0;
	const bool ok = loopback_capture_pid(o.listen, o.seconds, pcm, rate, channels, bits, tag, block);
	int rc = 1;
	if (ok && !pcm.empty()) {
		std::wstring out = o.wav_out.empty() ? (coop::exe_directory() + L"coop_listen.wav") : o.wav_out;
		if (coop::wav_write(out, pcm.data(), pcm.size(), rate, channels, bits, tag)) {
			std::wprintf(L"Wrote captured output: %ls\n", out.c_str());
		}
		// Trim the first ~0.7 s for analysis (loopback capture ramp-up) -- the .wav keeps it all.
		const std::size_t skip = std::min<std::size_t>(pcm.size(), static_cast<std::size_t>(rate) * block * 7 / 10);
		auto mono = coop::decode_channel(pcm.data() + skip, pcm.size() - skip, tag, bits, channels, 0);
		if (!mono.empty()) {
			std::printf("\n[LISTEN] pid %lu render output (what the guest hears):\n", o.listen);
			const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), rate, o.freq);
			print_report(r, o.freq, rate, UINT64_MAX);
			rc = 0;
		}
	} else {
		std::printf("ERROR: no audio captured from pid %lu (is it rendering?).\n", o.listen);
	}
	if (com_ok) {
		CoUninitialize();
	}
	return rc;
}

// --selfcheck: render a clean sine IN THIS PROCESS (no mirror) and self-loopback-capture
// it. The control for --render: it shares the exact self-capture path but with a known-good
// renderer, so if it reads clean (~60 dB, no gaps) then any defect --render shows is the
// mirror's, not an artifact of capturing our own process.
int run_selfcheck_mode(const Options& o)
{
	const bool com_ok = SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED));
	std::atomic<bool> stop{false};
	std::thread renderer([&]() {
		if (FAILED(CoInitializeEx(nullptr, COINIT_MULTITHREADED))) {
			return;
		}
		coop::tone::ToneSource tone;
		coop::tone::ToneFormat tf; // {} = device mix format (no resample), cleanest reference
		if (tone.open(tf, o.freq)) {
			while (!stop.load(std::memory_order_relaxed)) {
				tone.render_step(100);
			}
			tone.close();
		}
		CoUninitialize();
	});
	Sleep(500); // let the in-process tone reach steady state

	std::printf("Self-rendering a clean %.0f Hz tone in-process + self-capturing (control)...\n", o.freq);
	std::vector<std::uint8_t> pcm;
	std::uint32_t rate = 0, channels = 0, bits = 0, tag = 0, block = 0;
	const bool ok = loopback_capture_pid(GetCurrentProcessId(), o.seconds, pcm, rate, channels, bits, tag, block);
	stop.store(true, std::memory_order_relaxed);
	renderer.join();

	int rc = 1;
	if (ok && !pcm.empty()) {
		const std::size_t skip = std::min<std::size_t>(pcm.size(), static_cast<std::size_t>(rate) * block * 7 / 10);
		auto mono = coop::decode_channel(pcm.data() + skip, pcm.size() - skip, tag, bits, channels, 0);
		if (!mono.empty()) {
			std::printf("\n[SELFCHECK] in-process tone via the self-capture path (control):\n");
			const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), rate, o.freq);
			print_report(r, o.freq, rate, UINT64_MAX);
			rc = 0;
		}
	} else {
		std::printf("ERROR: selfcheck produced no audio.\n");
	}
	if (com_ok) {
		CoUninitialize();
	}
	return rc;
}

// --baseline: loopback-capture the tone process DIRECTLY -- no hook, no mirror. This is the
// fidelity floor of the measurement chain itself (the tone's own AUTOCONVERTPCM render + the
// process-loopback capture). Comparing --render against this floor separates a real mirror
// defect from the measurement's own noise.
int run_baseline_mode(const Options& o, HANDLE tone_proc, unsigned long target_pid)
{
	const bool com_ok = SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED));
	std::printf("Loopback-capturing the tone directly (no hook, no mirror) -- measurement floor...\n");
	std::vector<std::uint8_t> pcm;
	std::uint32_t rate = 0, channels = 0, bits = 0, tag = 0, block = 0;
	const bool ok = loopback_capture_pid(target_pid, o.seconds, pcm, rate, channels, bits, tag, block);
	int rc = 1;
	if (ok && !pcm.empty()) {
		const std::size_t skip = std::min<std::size_t>(pcm.size(), static_cast<std::size_t>(rate) * block * 7 / 10);
		auto mono = coop::decode_channel(pcm.data() + skip, pcm.size() - skip, tag, bits, channels, 0);
		if (!mono.empty()) {
			std::printf("\n[BASELINE] tone direct (measurement floor):\n");
			const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), rate, o.freq);
			print_report(r, o.freq, rate, UINT64_MAX);
			rc = 0;
		}
	} else {
		std::printf("ERROR: baseline loopback capture produced no audio.\n");
	}
	if (tone_proc != nullptr) {
		TerminateProcess(tone_proc, 0);
		CloseHandle(tone_proc);
	}
	if (com_ok) {
		CoUninitialize();
	}
	return rc;
}

// --render: measure the host's RENDER path, not just the capture ring. We drive the REAL
// shipping AudioMirror (its run_hooked re-renders the captured ring to the output device,
// silencing the game), and at the same time loopback-capture THIS process's own audio --
// which is exactly what AudioMirror renders, *including* any under-run silence gaps the
// device actually played. That makes the choppy / "metallic" re-prime artifact visible
// (a write-side tap would miss it: the gap is silence the device inserts, not bytes we wrote).
int run_render_mode(const Options& o, HANDLE tone_proc, unsigned long target_pid)
{
	const bool com_ok = SUCCEEDED(CoInitializeEx(nullptr, COINIT_MULTITHREADED));

	WAVEFORMATEX* mix = coop::default_render_format();
	if (mix == nullptr) {
		std::printf("ERROR: could not get the default render format.\n");
		if (com_ok) {
			CoUninitialize();
		}
		return 1;
	}
	std::uint32_t rate = 0, channels = 0, bits = 0, tag = 0;
	resolve_waveformat(mix, rate, channels, bits, tag);
	std::printf("Render endpoint mix format: %u Hz / %u ch / %u-bit / tag %u\n", rate, channels, bits, tag);

	// Capture our own render output (the mirror's). Game audio is silenced on the hooked
	// path, so our process's render mix == exactly what the guest would hear.
	std::vector<std::uint8_t> rendered;
	std::mutex rendered_mutex;
	coop::ProcessLoopbackCapture selfcap;
	const bool cap_ok =
	    selfcap.start(GetCurrentProcessId(), mix, [&](const BYTE* data, std::uint32_t frames, bool silent) {
		    const std::size_t bytes = static_cast<std::size_t>(frames) * mix->nBlockAlign;
		    std::lock_guard<std::mutex> lk(rendered_mutex);
		    if (silent || data == nullptr) {
			    rendered.insert(rendered.end(), bytes, 0);
		    } else {
			    rendered.insert(rendered.end(), data, data + bytes);
		    }
	    });
	if (!cap_ok) {
		std::printf("ERROR: self-loopback capture failed to start.\n");
		CoTaskMemFree(mix);
		if (com_ok) {
			CoUninitialize();
		}
		return 1;
	}

	// Drive the real mirror: it discovers the hook's ring, re-renders it (silencing the game).
	coop::AudioMirror mirror;
	if (!mirror.start(target_pid)) {
		std::printf("ERROR: AudioMirror failed to start.\n");
	}
	std::printf("Rendering through the real AudioMirror for %d s (source warms up, then measure)...\n", o.seconds);
	Sleep(static_cast<DWORD>(o.seconds) * 1000);
	std::printf("  mirror: source=%s status=\"%s\" buffered=%u ms\n", mirror.source_name(), mirror.status().c_str(),
	            mirror.buffered_ms());
	mirror.stop();
	selfcap.stop();

	std::vector<std::uint8_t> pcm;
	{
		std::lock_guard<std::mutex> lk(rendered_mutex);
		pcm.swap(rendered);
	}

	// Trim the first ~0.7 s: it contains start-up priming / the loopback warming up, which
	// would otherwise read as a spurious leading dropout.
	const std::size_t skip =
	    std::min<std::size_t>(pcm.size(), static_cast<std::size_t>(rate) * mix->nBlockAlign * 7 / 10);
	const std::uint8_t* body = pcm.data() + skip;
	const std::size_t body_bytes = pcm.size() - skip;

	std::wstring out = o.wav_out.empty() ? (coop::exe_directory() + L"coop_render.wav") : o.wav_out;
	if (coop::wav_write(out, body, body_bytes, rate, channels, bits, tag)) {
		std::wprintf(L"Wrote rendered output: %ls\n", out.c_str());
	}

	auto mono = coop::decode_channel(body, body_bytes, tag, bits, channels, 0);
	if (mono.empty()) {
		std::printf("NOTE: render format isn't float32/int16; WAV written, analysis skipped.\n");
	} else {
		std::printf("\n[RENDER PATH] what the guest actually hears (real AudioMirror output):\n");
		const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), rate, o.freq);
		print_report(r, o.freq, rate, UINT64_MAX);
	}

	CoTaskMemFree(mix);
	if (tone_proc != nullptr) {
		TerminateProcess(tone_proc, 0);
		CloseHandle(tone_proc);
	}
	if (com_ok) {
		CoUninitialize();
	}
	return mono.empty() ? 1 : 0;
}

bool parse_args(int argc, wchar_t** argv, Options& o)
{
	for (int i = 1; i < argc; ++i) {
		const std::wstring a = argv[i];
		auto next = [&](unsigned& dst) {
			if (i + 1 < argc) {
				dst = static_cast<unsigned>(_wtoi(argv[++i]));
			}
		};
		if (a == L"--pid" && i + 1 < argc) {
			o.pid = std::wcstoul(argv[++i], nullptr, 10);
		} else if (a == L"--listen" && i + 1 < argc) {
			o.listen = std::wcstoul(argv[++i], nullptr, 10);
		} else if (a == L"--freq" && i + 1 < argc) {
			o.freq = _wtof(argv[++i]);
		} else if (a == L"--rate") {
			next(o.rate);
		} else if (a == L"--channels") {
			next(o.channels);
		} else if (a == L"--bits") {
			next(o.bits);
		} else if (a == L"--seconds" && i + 1 < argc) {
			o.seconds = std::max(1, _wtoi(argv[++i]));
		} else if (a == L"--render") {
			o.render = true;
		} else if (a == L"--baseline") {
			o.baseline = true;
		} else if (a == L"--selfcheck") {
			o.selfcheck = true;
		} else if (a == L"--wav" && i + 1 < argc) {
			o.wav_in = argv[++i];
		} else if (a == L"--out" && i + 1 < argc) {
			o.wav_out = argv[++i];
		} else if (a == L"--help" || a == L"-h") {
			return false;
		}
	}
	return true;
}

} // namespace

int wmain(int argc, wchar_t** argv)
{
	Options o;
	if (!parse_args(argc, argv, o)) {
		std::printf("usage: coop_audio_validate [--pid N] [--listen N] [--freq Hz] [--rate Hz]\n"
		            "       [--channels N] [--bits 16|32] [--seconds N] [--render | --baseline | --selfcheck]\n"
		            "       [--wav file] [--out file]\n"
		            "  (no args)  spawn coop_tone @ 44100/1000 Hz, capture the hook ring, analyze.\n"
		            "  --render   also drive the real AudioMirror and measure its rendered output\n"
		            "             (surfaces under-run / re-prime 'metallic' gaps the capture side can't show).\n"
		            "  --baseline loopback-capture the tone directly (no hook/mirror) = the measurement floor.\n"
		            "  --selfcheck render a clean tone in-process + self-capture (control for the self-capture path).\n"
		            "  --listen N passively record pid N's output to a .wav and analyze it -- point it at the\n"
		            "             live coop_host to hear/quantify exactly what the guest gets on a real game.\n"
		            "  --wav F    just analyze a recorded .wav.\n");
		return 1;
	}

	// --- Mode C: analyze a recorded .wav ---------------------------------------------
	if (!o.wav_in.empty()) {
		coop::WavData wd;
		if (!coop::wav_read(o.wav_in, wd)) {
			std::wprintf(L"ERROR: could not read WAV '%ls'.\n", o.wav_in.c_str());
			return 1;
		}
		std::printf("Loaded WAV: %u Hz / %u ch / %u-bit / tag %u, %zu bytes\n", wd.sample_rate, wd.channels, wd.bits,
		            wd.format_tag, wd.pcm.size());
		auto mono = coop::decode_channel(wd.pcm.data(), wd.pcm.size(), wd.format_tag, wd.bits, wd.channels, 0);
		if (mono.empty()) {
			std::printf("ERROR: unsupported WAV sample format (need 16-bit PCM or 32-bit float).\n");
			return 1;
		}
		const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), wd.sample_rate, o.freq);
		print_report(r, o.freq, wd.sample_rate, UINT64_MAX);
		return 0;
	}

	// --- Control: render a clean tone in-process + self-capture (no target needed) ----
	if (o.selfcheck) {
		return run_selfcheck_mode(o);
	}

	// --- Live: passively record an already-running process's output (e.g. coop_host) --
	if (o.listen != 0) {
		return run_listen_mode(o);
	}

	// --- Acquire a target: spawn coop_tone, or attach to a given pid ------------------
	HANDLE tone_proc = nullptr;
	unsigned long target_pid = o.pid;
	if (target_pid == 0) {
		tone_proc = spawn_tone(o, target_pid);
		if (tone_proc == nullptr) {
			return 1;
		}
		Sleep(700); // let the tone reach steady state before we inject
	} else {
		std::printf("Attaching to existing pid %lu (tone freq assumed %.0f Hz).\n", target_pid, o.freq);
	}

	// --- Mode: measurement floor (no hook, no mirror) ---------------------------------
	if (o.baseline) {
		return run_baseline_mode(o, tone_proc, target_pid);
	}

	// --- Set up the IPC the hook expects, then inject (late attach: ring AFTER inject) -
	coop::SharedMemory ipc;
	if (!ipc.create(coop::shared_memory_name(target_pid), sizeof(coop::SharedBlock))) {
		std::printf("ERROR: create input mapping failed (%lu).\n", GetLastError());
		return 1;
	}
	auto* block = ipc.as<coop::SharedBlock>();
	block->version = coop::kProtocolVersion;
	block->pad_count = 0;
	block->sequence.store(0, std::memory_order_relaxed);
	block->magic = coop::kProtocolMagic;

	enable_hook_trace();

	std::printf("Injecting coop_hook.dll into pid %lu ...\n", target_pid);
	if (!inject(target_pid, coop::deployed_artifact_path(L"coop_hook.dll"))) {
		std::printf("ERROR: injection failed.\n");
		return 1;
	}

	// --- Mode B: measure the host RENDER path (real AudioMirror) ----------------------
	if (o.render) {
		Sleep(1200); // let the hook register the stream before the mirror reads it
		const int rc = run_render_mode(o, tone_proc, target_pid);
		block->magic = 0;
		return rc;
	}

	// Create the audio ring ~1.5 s after injection -- this is the real app's ordering (the
	// host creates the ring only when audio mirroring is toggled on), and it forces the
	// hook's late-attach guess+measure path (the exact Brotato scenario).
	Sleep(1500);
	coop::SharedMemory ring_shm;
	if (!ring_shm.create(coop::audio_ring_name(target_pid), coop::audio_ring_total_size(coop::kAudioRingCapacity))) {
		std::printf("ERROR: create audio ring mapping failed (%lu).\n", GetLastError());
		return 1;
	}
	auto* ring = ring_shm.as<coop::AudioRingHeader>();
	coop::audio_ring_init(*ring, coop::kAudioRingCapacity);
	ring->capture_enabled.store(1, std::memory_order_release);

	// --- Capture + analyze ------------------------------------------------------------
	std::vector<std::uint8_t> pcm;
	std::uint32_t rate = 0, channels = 0, bits = 0, format_tag = 0;
	std::uint64_t overruns = 0;
	const bool captured = capture_ring(ring, o.seconds, pcm, rate, channels, bits, format_tag, overruns);

	if (captured) {
		// Dump the captured audio so it can be listened to.
		std::wstring out = o.wav_out.empty() ? (coop::exe_directory() + L"coop_capture.wav") : o.wav_out;
		if (coop::wav_write(out, pcm.data(), pcm.size(), rate, channels, bits, format_tag)) {
			std::wprintf(L"Wrote captured audio: %ls\n", out.c_str());
		}

		auto mono = coop::decode_channel(pcm.data(), pcm.size(), format_tag, bits, channels, 0);
		if (mono.empty()) {
			std::printf("NOTE: captured format isn't float32/int16, can't decode for analysis (WAV still written).\n");
		} else {
			const coop::ToneReport r = coop::analyze_tone(mono.data(), mono.size(), rate, o.freq);
			print_report(r, o.freq, rate, overruns);
		}
	}

	block->magic = 0; // invalidate so a late hook read won't trust stale data
	if (tone_proc != nullptr) {
		TerminateProcess(tone_proc, 0);
		CloseHandle(tone_proc);
	}
	return captured ? 0 : 1;
}