From 62097c3e5ea431e2f287264870dde82e905c3679 Mon Sep 17 00:00:00 2001
From: Stuart Cameron
Date: Thu, 19 Nov 2026 20:50:37 +1100
Subject: [PATCH 1/5] ci(task #72): name vspipe's teardown race in crash
reports; keep the raw .ips
print-crash-reports.py prints every thread of a macOS crash report (the
race's other half is main, freeing the script), classifies each crash by the
first frame outside the system libraries, and names task #68's teardown race
(vspipe's frameDoneCallback/outputFrame) together with whether the installed
vspipe carries the fix. --classify FILE... prints one verdict per .ips or gdb
text. The macOS jobs of ci-test and nightly upload the raw .ips on failure.
crash_report_classifier_test.dart (push gate) checks it on the crashes CI
printed: the macOS x64 SIGILL (os_unfair_lock abort under notify_one), the
Linux SIGSEGV (outputFrame, main in vpy4_freeScript), and a MoltenVK crash
that is not the race.
Co-Authored-By: Claude Opus 5.5
---
.github/scripts/print-crash-reports.py | 223 +++++++++++++++---
.github/workflows/ci-test.yml | 20 ++
.github/workflows/nightly.yml | 20 ++
Tests/TestResources/crash_reports/README.md | 19 ++
.../linux-vspipe-teardown-sigsegv.gdb.txt | 81 +++++++
.../macos-arm64-moltenvk-sigsegv.ips | 73 ++++++
.../macos-x64-vspipe-teardown-sigill.ips | 76 ++++++
app/test/crash_report_classifier_test.dart | 65 +++++
8 files changed, 548 insertions(+), 29 deletions(-)
create mode 100644 Tests/TestResources/crash_reports/README.md
create mode 100644 Tests/TestResources/crash_reports/linux-vspipe-teardown-sigsegv.gdb.txt
create mode 100644 Tests/TestResources/crash_reports/macos-arm64-moltenvk-sigsegv.ips
create mode 100644 Tests/TestResources/crash_reports/macos-x64-vspipe-teardown-sigill.ips
create mode 100644 app/test/crash_report_classifier_test.dart
diff --git a/.github/scripts/print-crash-reports.py b/.github/scripts/print-crash-reports.py
index 62c3f55f..ed054442 100644
--- a/.github/scripts/print-crash-reports.py
+++ b/.github/scripts/print-crash-reports.py
@@ -6,7 +6,9 @@
- macOS writes a crash report (.ips) for every crashed process, naming the
library and function that faulted; this prints the newest few for the given
- process names.
+ process names: the faulting thread, then every other thread's top frames
+ (task #72: the teardown race's other half is the main thread, in
+ `freeScript` or `exit`). The workflows also upload the raw reports.
- Linux keeps nothing unless the job asked for core dumps
(`.github/scripts/enable-core-dumps.sh`: `kernel.core_pattern` piped to a
handler that writes into `$VAPOURBOX_CORE_DIR`, default /tmp/cores). This
@@ -15,9 +17,21 @@
frame can still be resolved later (task #66).
- Windows: nothing to read; says so.
+Each crash is then classified: the first frame outside the system libraries
+names the image that faulted, and one signature is recognised by name,
+vspipe R78's teardown race (task #68): that frame is vspipe's own
+`frameDoneCallback` or `outputFrame`. On macOS it arrives as SIGILL
+(`pthread_cond_signal` on the freed condition variable traps in
+`_os_unfair_lock_unowned_abort`), on Linux as SIGSEGV (task #72). The line
+also says whether the installed bundle's vspipe carries the fix (its
+`--version` marker, read from the binary).
+
Never fails the step: it is diagnostic.
-Usage: print-crash-reports.py [process-name ...] (default: vspipe-bin python3)
+Usage:
+ print-crash-reports.py [process-name ...] (default: vspipe-bin python3 python3.12)
+ print-crash-reports.py --classify FILE ... one line per file: FILEverdict
+ (.ips, or gdb backtrace text)
"""
import glob
import json
@@ -27,42 +41,171 @@
import subprocess
import sys
-NAMES = sys.argv[1:] or ["vspipe-bin", "python3", "python3.12"]
+DEFAULT_NAMES = ["vspipe-bin", "python3", "python3.12"]
NEWEST = 3
+TEARDOWN = "vspipe-teardown"
+# The marker task #68's patch adds to `vspipe --version`; it is a literal in
+# the binary, so it can be found without running vspipe (which needs the
+# bundle's environment).
+FIX_MARKER = b"VapourBox patch: vspipe frame-done count"
+VSPIPE_IMAGES = {"vspipe", "vspipe-bin", "vspipe-bin.real", "vspipe.exe"}
+VSPIPE_CALLBACKS = ("frameDoneCallback", "outputFrame")
+TEARDOWN_MAIN = ("freeScript", "vpy4_freeScript", "exit", "__cxa_finalize")
+_SYSTEM = re.compile(
+ r"^(libsystem_\w+\.dylib|libc\+\+(abi)?\.1\.dylib|libdyld\.dylib|dyld"
+ r"|libc\.so\.\d+|libpthread\.so\.\d+|libstdc\+\+\.so\.\d+|libgcc_s\.so\.\d+"
+ r"|ld-linux[-\w]*\.so\.\d+|\?\?)$"
+)
+
+
+def is_system(image):
+ return bool(_SYSTEM.match(image or "??"))
+
+
+def classify(frames, other_threads=()):
+ """frames: the faulting thread, top first, as (image basename, symbol).
+ other_threads: the other threads' frames, the same way. Returns
+ (verdict, detail): verdict is TEARDOWN for task #68's race, else
+ "image:" for the first frame outside the system libraries, else
+ "unknown"."""
+ first = next(((img, sym) for img, sym in frames if not is_system(img)), None)
+ if first is None:
+ return "unknown", "every frame is in a system library"
+ img, sym = first
+ if img in VSPIPE_IMAGES and any(cb in sym for cb in VSPIPE_CALLBACKS):
+ main = next((s for t in other_threads for _, s in t[:12]
+ if any(m in s for m in TEARDOWN_MAIN)), None)
+ detail = f"{img} {sym.split('(')[0]}"
+ if main:
+ detail += f"; another thread is in {main.split('(')[0]}"
+ return TEARDOWN, detail
+ return f"image:{img}", sym
+
+
+def vspipe_fix_state():
+ """'yes' / 'no' for the installed bundle's vspipe, or None if not found."""
+ deps = os.environ.get("VAPOURBOX_DEPS_DIR", "")
+ if not deps:
+ return None
+ for pattern in ("vapoursynth/vspipe-bin.real", "vapoursynth/vspipe-bin",
+ "vapoursynth/Lib/site-packages/vapoursynth/vspipe.exe"):
+ path = os.path.join(deps, pattern)
+ if os.path.isfile(path):
+ with open(path, "rb") as f:
+ return "yes" if FIX_MARKER in f.read() else "no"
+ return None
+
+
+def verdict_line(verdict, detail):
+ if verdict == TEARDOWN:
+ fix = vspipe_fix_state()
+ has = {"yes": "the installed vspipe carries the fix, so this is NOT that race being unfixed: look again",
+ "no": "the installed vspipe does not carry the fix (deps before 1.18.0)",
+ None: "could not check the installed vspipe for the fix"}[fix]
+ return (f" => vspipe R78's teardown race (task #68): {detail}. "
+ f"Not a plugin or an instruction-set fault; {has}.")
+ if verdict.startswith("image:"):
+ return f" => faulted in {verdict[6:]}: {detail}"
+ return f" => {detail}"
+
+
+# --- macOS -----------------------------------------------------------------
+
+def ips_threads(path):
+ """(exception summary, faulting thread frames, other threads' frames)."""
+ with open(path, encoding="utf-8", errors="replace") as f:
+ f.readline() # one-line JSON header
+ body = json.loads(f.read())
+ exc = body.get("exception", {})
+ summary = f"{exc.get('type')} {exc.get('signal')} {exc.get('subtype', '')}".rstrip()
+ images = body.get("usedImages", [])
+
+ def frames_of(thread):
+ out = []
+ for frame in thread.get("frames", []):
+ idx = frame.get("imageIndex", -1)
+ image = images[idx] if 0 <= idx < len(images) else {}
+ lib = os.path.basename(image.get("path", image.get("name", "??")) or "??")
+ sym = frame.get("symbol", f"+{frame.get('imageOffset', 0):#x}")
+ out.append((lib, sym))
+ return out
+
+ threads = body.get("threads", [])
+ faulting = body.get("faultingThread", 0)
+ fault = frames_of(threads[faulting]) if faulting < len(threads) else []
+ others = [(i, frames_of(t)) for i, t in enumerate(threads) if i != faulting]
+ return summary, fault, others
+
+
+def print_ips(path):
+ print(f"=== {os.path.basename(path)}")
+ try:
+ summary, fault, others = ips_threads(path)
+ except (OSError, ValueError) as e:
+ print(f" (unparseable: {e})")
+ return
+ print(f" {summary}")
+ print(" --- faulting thread")
+ for i, (lib, sym) in enumerate(fault[:30]):
+ print(f" {i:2} {lib:32} {sym}")
+ for n, frames in others:
+ print(f" --- thread {n}")
+ for i, (lib, sym) in enumerate(frames[:8]):
+ print(f" {i:2} {lib:32} {sym}")
+ print(verdict_line(*classify(fault, [f for _, f in others])))
-def macos():
+
+def macos(names):
dirs = [os.path.expanduser("~/Library/Logs/DiagnosticReports"), "/Library/Logs/DiagnosticReports"]
reports = sorted(
(p for d in dirs for p in glob.glob(os.path.join(d, "*.ips"))
- if any(os.path.basename(p).startswith(n) for n in NAMES)),
+ if any(os.path.basename(p).startswith(n) for n in names)),
key=os.path.getmtime,
reverse=True,
)[:NEWEST]
if not reports:
- print(f"No crash reports for {', '.join(NAMES)}.")
+ print(f"No crash reports for {', '.join(names)}.")
for path in reports:
- print(f"=== {os.path.basename(path)}")
- with open(path, encoding="utf-8", errors="replace") as f:
- f.readline() # one-line JSON header
- try:
- body = json.loads(f.read())
- except json.JSONDecodeError as e:
- print(f" (unparseable: {e})")
- continue
- exc = body.get("exception", {})
- print(f" {exc.get('type')} {exc.get('signal')} {exc.get('subtype', '')}".rstrip())
- images = body.get("usedImages", [])
- threads = body.get("threads", [])
- faulting = body.get("faultingThread", 0)
- if faulting >= len(threads):
- continue
- for i, frame in enumerate(threads[faulting].get("frames", [])[:30]):
- image = images[frame.get("imageIndex", -1)] if 0 <= frame.get("imageIndex", -1) < len(images) else {}
- lib = os.path.basename(image.get("path", image.get("name", "?")))
- sym = frame.get("symbol", f"+{frame.get('imageOffset', 0):#x}")
- print(f" {i:2} {lib:32} {sym}")
+ print_ips(path)
+
+
+# --- Linux -----------------------------------------------------------------
+
+_GDB_FRAME = re.compile(r"^\s*#\d+\s+(?:0x[0-9a-fA-F]+ in )?(.*?)(?: from (\S+))?\s*$")
+
+
+def gdb_threads(text, exe="vspipe-bin"):
+ """(faulting thread frames, other threads' frames) from gdb's output.
+ The faulting thread is the block after "--- faulting thread" if this
+ script printed it, else the first backtrace in the text."""
+ def parse(block):
+ out = []
+ for line in block:
+ m = _GDB_FRAME.match(line)
+ if m:
+ sym, lib = m.group(1), m.group(2)
+ out.append((os.path.basename(lib) if lib else exe, sym))
+ return out
+
+ lines = text.splitlines()
+ blocks, cur, start = [], [], 0 # (index of the block's first line, its lines)
+ for i, line in enumerate(lines):
+ if _GDB_FRAME.match(line) and line.strip().startswith("#"):
+ if not cur:
+ start = i
+ cur.append(line)
+ elif cur:
+ blocks.append((start, cur))
+ cur = []
+ if cur:
+ blocks.append((start, cur))
+ marker = next((i for i, l in enumerate(lines) if "--- faulting thread" in l), -1)
+ chosen = next((b for s, b in blocks if s > marker), None)
+ fault = parse(chosen) if chosen else []
+ others = [parse(b) for _, b in blocks if b is not chosen]
+ return fault, others
def core_executable(core):
@@ -124,12 +267,34 @@ def linux():
print(f" executable: {exe}")
for line in lines:
print(f" {line}")
+ fault, others = gdb_threads(text, os.path.basename(exe))
+ print(verdict_line(*classify(fault, others)))
+
+# --- command line ----------------------------------------------------------
-def main():
+def classify_file(path):
+ if path.endswith(".ips"):
+ _, fault, others = ips_threads(path)
+ return classify(fault, [f for _, f in others])
+ with open(path, encoding="utf-8", errors="replace") as f:
+ fault, others = gdb_threads(f.read())
+ return classify(fault, others)
+
+
+def main(argv):
+ if argv and argv[0] == "--classify":
+ for path in argv[1:]:
+ try:
+ verdict, detail = classify_file(path)
+ except (OSError, ValueError) as e:
+ verdict, detail = "unreadable", str(e)
+ print(f"{path}\t{verdict}\t{detail}")
+ return
+ names = argv or DEFAULT_NAMES
try:
if sys.platform == "darwin":
- macos()
+ macos(names)
elif sys.platform.startswith("linux"):
linux()
else:
@@ -139,4 +304,4 @@ def main():
if __name__ == "__main__":
- main()
+ main(sys.argv[1:])
diff --git a/.github/workflows/ci-test.yml b/.github/workflows/ci-test.yml
index 719f9287..f445961c 100644
--- a/.github/workflows/ci-test.yml
+++ b/.github/workflows/ci-test.yml
@@ -130,6 +130,26 @@ jobs:
if: failure()
run: python3 .github/scripts/print-crash-reports.py
+ - name: Keep the raw crash reports (on failure)
+ # The .ips files themselves, for symbolication or a second look: the
+ # step above prints only their top frames (task #72).
+ if: failure()
+ run: |
+ mkdir -p crash-reports
+ for f in ~/Library/Logs/DiagnosticReports/{vspipe-bin,python3,vapourbox-worker}*.ips; do
+ [ -f "$f" ] && cp "$f" crash-reports/
+ done
+ ls crash-reports
+
+ - name: Upload the raw crash reports (on failure)
+ if: failure()
+ uses: actions/upload-artifact@v5
+ with:
+ name: crash-reports-macos-${{ matrix.arch }}
+ path: crash-reports/*.ips
+ if-no-files-found: ignore
+ retention-days: 14
+
windows:
runs-on: windows-latest
env:
diff --git a/.github/workflows/nightly.yml b/.github/workflows/nightly.yml
index 3056351e..77f2575a 100644
--- a/.github/workflows/nightly.yml
+++ b/.github/workflows/nightly.yml
@@ -157,6 +157,26 @@ jobs:
if: failure()
run: python3 .github/scripts/print-crash-reports.py
+ - name: Keep the raw crash reports (on failure)
+ # The .ips files themselves, for symbolication or a second look: the
+ # step above prints only their top frames (task #72).
+ if: failure()
+ run: |
+ mkdir -p crash-reports
+ for f in ~/Library/Logs/DiagnosticReports/{vspipe-bin,python3,vapourbox-worker}*.ips; do
+ [ -f "$f" ] && cp "$f" crash-reports/
+ done
+ ls crash-reports
+
+ - name: Upload the raw crash reports (on failure)
+ if: failure()
+ uses: actions/upload-artifact@v5
+ with:
+ name: crash-reports-${{ matrix.asset }}-${{ matrix.suite }}
+ path: crash-reports/*.ips
+ if-no-files-found: ignore
+ retention-days: 14
+
windows:
if: ${{ github.event_name == 'schedule' || inputs.platform == 'all' || inputs.platform == 'windows' }}
# Both CPU-tier bundles (issue #92); see the macOS job for what the v2 run proves.
diff --git a/Tests/TestResources/crash_reports/README.md b/Tests/TestResources/crash_reports/README.md
new file mode 100644
index 00000000..38b0a79c
--- /dev/null
+++ b/Tests/TestResources/crash_reports/README.md
@@ -0,0 +1,19 @@
+# Crash-report fixtures (task #72)
+
+Inputs for `app/test/crash_report_classifier_test.dart`, which runs
+`.github/scripts/print-crash-reports.py --classify` over them.
+
+- `macos-x64-vspipe-teardown-sigill.ips`: the faulting thread of
+ `vspipe-bin-2026-11-18-023804.ips` from the push gate's macOS x64 job,
+ run 44088351262 attempt 2 (deps 1.17.0-rc3), rebuilt as an `.ips` from the
+ frames CI printed. CI printed only the faulting thread, so it has only that
+ thread; the image paths are placeholders. vspipe R78's teardown race (task
+ #68) as macOS reports it: SIGILL in `_os_unfair_lock_unowned_abort`.
+- `macos-arm64-moltenvk-sigsegv.ips`: run 44064242986's macOS arm64
+ `vspipe-bin-2026-11-17-103557.ips` (RIFE creating its Vulkan device
+ through MoltenVK), rebuilt the same way, with some frames left out. A
+ crash that is not the race.
+- `linux-vspipe-teardown-sigsegv.gdb.txt`: the gdb output CI printed for
+ `core.vspipe-bin.52015.1795010338`, run 44318045317's Linux job, from
+ "faulting thread" to the end of "every thread (top frames)". The same race
+ on Linux: SIGSEGV in `outputFrame`, main in `vpy4_freeScript`.
diff --git a/Tests/TestResources/crash_reports/linux-vspipe-teardown-sigsegv.gdb.txt b/Tests/TestResources/crash_reports/linux-vspipe-teardown-sigsegv.gdb.txt
new file mode 100644
index 00000000..f771ec2a
--- /dev/null
+++ b/Tests/TestResources/crash_reports/linux-vspipe-teardown-sigsegv.gdb.txt
@@ -0,0 +1,81 @@
+[New LWP 52133]
+[New LWP 52136]
+[New LWP 52135]
+[New LWP 52134]
+[New LWP 52015]
+[Thread debugging using libthread_db enabled]
+Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
+Core was generated by `/home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/vspipe-bin -c y4'.
+Program terminated with signal SIGSEGV, Segmentation fault.
+#0 0x000055c366d8e04b in outputFrame(VSFrame const*, VSPipeOutputData*) ()
+[Current thread is 1 (Thread 0x7fac173fe6c0 (LWP 52133))]
+
+--- faulting thread
+#0 0x000055c366d8e04b in outputFrame(VSFrame const*, VSPipeOutputData*) ()
+#1 0x000055c366d8eaad in frameDoneCallback(void*, VSFrame const*, int, VSNode*, char const*) [clone .lto_priv.0] ()
+#2 0x00007fac1e92ef8b in VSThreadPool::returnFrame(VSFrameContext*, vs_intrusive_ptr const&, std::unique_lock&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#3 0x00007fac1e931b25 in VSThreadPool::runTasks(bool&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#4 0x00007fac20eecdb4 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
+#5 0x00007fac20a9cb84 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#6 0x00007fac20b29ecc in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+
+--- registers
+rip 0x55c366d8e04b 0x55c366d8e04b
+rsp 0x7fac173fd560 0x7fac173fd560
+rbp 0x7fac100bbdb0 0x7fac100bbdb0
+rdi 0x7fac100bbdb0 140376980307376
+rsi 0x1 1
+rdx 0x1 1
+rcx 0x7fac20a99098 140377259085976
+=> 0x55c366d8e04b <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+587>:
+ 0x55c366d8e051 <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+593>:
+ 0x55c366d8e054 <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+596>:
+ 0x55c366d8e057 <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+599>:
+ 0x55c366d8e05a <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+602>:
+ 0x55c366d8e05d <_ZL11outputFramePK7VSFrameP16VSPipeOutputData+605>:
+
+--- every thread (top frames)
+
+Thread 5 (Thread 0x7fac21152740 (LWP 52015)):
+#0 0x00007fac1f44432b in visit_reachable () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#1 0x00007fac1f39560e in dict_traverse () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#2 0x00007fac1f444277 in deduce_unreachable () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#3 0x00007fac1f4427f5 in gc_collect_main.llvm () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#4 0x00007fac1f574275 in gc_collect () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#5 0x00007fac1f398876 in cfunction_vectorcall_FASTCALL_KEYWORDS.llvm.16519472471476773533 () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#6 0x00007fac1f36b486 in PyObject_Vectorcall () from /home/runner/work/VapourBox/VapourBox/deps/linux-x64/python/lib/libpython3.12.so.1.0
+#7 0x00007fac1eaaced2 in vpy4_freeScript(VSScript*) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/vapoursynth.abi3.so
+
+Thread 4 (Thread 0x7fac163fd6c0 (LWP 52134)):
+#0 0x00007fac20a98e51 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#1 0x00007fac20a9b8cd in pthread_cond_wait () from /lib/x86_64-linux-gnu/libc.so.6
+#2 0x00007fac1e930975 in VSThreadPool::runTasks(bool&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#3 0x00007fac20eecdb4 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
+#4 0x00007fac20a9cb84 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#5 0x00007fac20b29ecc in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+
+Thread 3 (Thread 0x7fac153fc6c0 (LWP 52135)):
+#0 0x00007fac20a98e51 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#1 0x00007fac20a9b8cd in pthread_cond_wait () from /lib/x86_64-linux-gnu/libc.so.6
+#2 0x00007fac1e930975 in VSThreadPool::runTasks(bool&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#3 0x00007fac20eecdb4 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
+#4 0x00007fac20a9cb84 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#5 0x00007fac20b29ecc in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+
+Thread 2 (Thread 0x7fac0ffff6c0 (LWP 52136)):
+#0 0x00007fac20a98e51 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#1 0x00007fac20a9b8cd in pthread_cond_wait () from /lib/x86_64-linux-gnu/libc.so.6
+#2 0x00007fac1e930975 in VSThreadPool::runTasks(bool&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#3 0x00007fac20eecdb4 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
+#4 0x00007fac20a9cb84 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#5 0x00007fac20b29ecc in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+
+Thread 1 (Thread 0x7fac173fe6c0 (LWP 52133)):
+#0 0x000055c366d8e04b in outputFrame(VSFrame const*, VSPipeOutputData*) ()
+#1 0x000055c366d8eaad in frameDoneCallback(void*, VSFrame const*, int, VSNode*, char const*) [clone .lto_priv.0] ()
+#2 0x00007fac1e92ef8b in VSThreadPool::returnFrame(VSFrameContext*, vs_intrusive_ptr const&, std::unique_lock&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#3 0x00007fac1e931b25 in VSThreadPool::runTasks(bool&) () from /home/runner/.local/share/VapourBox/deps/linux-x64/vapoursynth/libvapoursynth.so.4
+#4 0x00007fac20eecdb4 in ?? () from /lib/x86_64-linux-gnu/libstdc++.so.6
+#5 0x00007fac20a9cb84 in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+#6 0x00007fac20b29ecc in ?? () from /lib/x86_64-linux-gnu/libc.so.6
+
diff --git a/Tests/TestResources/crash_reports/macos-arm64-moltenvk-sigsegv.ips b/Tests/TestResources/crash_reports/macos-arm64-moltenvk-sigsegv.ips
new file mode 100644
index 00000000..3c7c147f
--- /dev/null
+++ b/Tests/TestResources/crash_reports/macos-arm64-moltenvk-sigsegv.ips
@@ -0,0 +1,73 @@
+{"app_name": "vspipe-bin", "bug_type": "309"}
+{
+ "exception": {
+ "type": "EXC_BAD_ACCESS",
+ "signal": "SIGSEGV",
+ "subtype": "KERN_INVALID_ADDRESS at 0x000000000000000c"
+ },
+ "faultingThread": 0,
+ "usedImages": [
+ {
+ "name": "AppleParavirtGPUMetalIOGPUFamily",
+ "path": "/x/AppleParavirtGPUMetalIOGPUFamily"
+ },
+ {
+ "name": "libMoltenVK.dylib",
+ "path": "/x/libMoltenVK.dylib"
+ },
+ {
+ "name": "librife.dylib",
+ "path": "/x/librife.dylib"
+ },
+ {
+ "name": "libvapoursynth.4.dylib",
+ "path": "/x/libvapoursynth.4.dylib"
+ },
+ {
+ "name": "vspipe-bin",
+ "path": "/x/vspipe-bin"
+ },
+ {
+ "name": "dyld",
+ "path": "/x/dyld"
+ }
+ ],
+ "threads": [
+ {
+ "frames": [
+ {
+ "imageIndex": 0,
+ "imageOffset": 27772
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "MVKImagePlane::updateMTLTextureContent(MVKImageSubresource&, unsigned long long, unsigned long long)"
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "vkBindImageMemory"
+ },
+ {
+ "imageIndex": 2,
+ "symbol": "ncnn::VkBlobAllocator::fastMalloc(int, int, int, unsigned long, int)"
+ },
+ {
+ "imageIndex": 2,
+ "symbol": "rifeCreate(VSMap const*, VSMap*, void*, VSCore*, VSAPI const*)"
+ },
+ {
+ "imageIndex": 3,
+ "symbol": "invoke(VSPlugin*, char const*, VSMap const*)"
+ },
+ {
+ "imageIndex": 4,
+ "symbol": "main"
+ },
+ {
+ "imageIndex": 5,
+ "symbol": "start"
+ }
+ ]
+ }
+ ]
+}
diff --git a/Tests/TestResources/crash_reports/macos-x64-vspipe-teardown-sigill.ips b/Tests/TestResources/crash_reports/macos-x64-vspipe-teardown-sigill.ips
new file mode 100644
index 00000000..f28a5c6c
--- /dev/null
+++ b/Tests/TestResources/crash_reports/macos-x64-vspipe-teardown-sigill.ips
@@ -0,0 +1,76 @@
+{"app_name": "vspipe-bin", "bug_type": "309"}
+{
+ "exception": {
+ "type": "EXC_BAD_INSTRUCTION",
+ "signal": "SIGILL"
+ },
+ "faultingThread": 0,
+ "usedImages": [
+ {
+ "name": "libsystem_platform.dylib",
+ "path": "/x/libsystem_platform.dylib"
+ },
+ {
+ "name": "libsystem_pthread.dylib",
+ "path": "/x/libsystem_pthread.dylib"
+ },
+ {
+ "name": "libc++.1.dylib",
+ "path": "/x/libc++.1.dylib"
+ },
+ {
+ "name": "vspipe-bin",
+ "path": "/x/vspipe-bin"
+ },
+ {
+ "name": "libvapoursynth.4.dylib",
+ "path": "/x/libvapoursynth.4.dylib"
+ }
+ ],
+ "threads": [
+ {
+ "frames": [
+ {
+ "imageIndex": 0,
+ "symbol": "_os_unfair_lock_unowned_abort"
+ },
+ {
+ "imageIndex": 0,
+ "symbol": "_os_unfair_lock_unlock_slow"
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "_pthread_cond_check_init_slow"
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "pthread_cond_signal"
+ },
+ {
+ "imageIndex": 2,
+ "symbol": "std::__1::condition_variable::notify_one()"
+ },
+ {
+ "imageIndex": 3,
+ "symbol": "frameDoneCallback(void*, VSFrame const*, int, VSNode*, char const*)"
+ },
+ {
+ "imageIndex": 4,
+ "symbol": "VSThreadPool::returnFrame(VSFrameContext*, vs_intrusive_ptr const&, std::__1::unique_lock&)"
+ },
+ {
+ "imageIndex": 4,
+ "symbol": "VSThreadPool::runTasks(bool&)"
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "_pthread_start"
+ },
+ {
+ "imageIndex": 1,
+ "symbol": "thread_start"
+ }
+ ]
+ }
+ ]
+}
diff --git a/app/test/crash_report_classifier_test.dart b/app/test/crash_report_classifier_test.dart
new file mode 100644
index 00000000..19552cbb
--- /dev/null
+++ b/app/test/crash_report_classifier_test.dart
@@ -0,0 +1,65 @@
+// The CI crash-report step names vspipe R78's teardown race (task #68) for
+// what it is, on both platforms, and names the faulting image of anything
+// else (task #72).
+//
+// The macOS x64 push gate failed a random format-sweep row in about half its
+// runs with "vspipe exited with signal 4 (SIGILL)", which reads like a plugin
+// using an instruction the runner lacks. It was the teardown race: vspipe's
+// frameDoneCallback signalling a condition variable main had already freed,
+// which macOS traps in _os_unfair_lock_unowned_abort (SIGILL); on Linux the
+// same race is a SIGSEGV in outputFrame. The fixtures are those crashes as
+// CI printed them (Tests/TestResources/crash_reports/README.md).
+import 'dart:io';
+
+import 'package:flutter_test/flutter_test.dart';
+import 'package:path/path.dart' as p;
+
+import 'support/worker_harness.dart';
+
+void main() {
+ final script = p.join(WorkerHarness.repoRoot, '.github', 'scripts', 'print-crash-reports.py');
+ final fixtures = p.join(WorkerHarness.repoRoot, 'Tests', 'TestResources', 'crash_reports');
+
+ // The system Python, as the workflows' crash-report step runs it; the script
+ // needs nothing but the standard library.
+ final python = Platform.isWindows ? 'python' : 'python3';
+
+ Future
+2026-11-19, measured in task #72 (run 44500126463,
+its report, section 3): with only vspipe-bin swapped in a
+1.17.0-rc3 bundle, R78 crashed 71 of 32,160 one-frame runs and the 1.18.0-rc1 vspipe 0 (Linux x64 66 vs 0, macOS x64 2 vs 0,
+Windows x64 3 vs 0), with this report's compare_vspipe.sh under CPU load. The workflow is kept in task #72's
+scripts/.
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-loops-linux-x64.txt b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-linux-x64.txt
new file mode 100644
index 00000000..9fdd073f
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-linux-x64.txt
@@ -0,0 +1,6 @@
+unpatched trigger (4 loops, CPU load): runs=4000 failures=0
+unpatched ivtc (4 loops, CPU load): runs=3360 failures=66
+unpatched bwdif (4 loops, CPU load): runs=3360 failures=0
+patched trigger (4 loops, CPU load): runs=4000 failures=0
+patched ivtc (4 loops, CPU load): runs=3360 failures=0
+patched bwdif (4 loops, CPU load): runs=3360 failures=0
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64-crash-reports.tsv b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64-crash-reports.tsv
new file mode 100644
index 00000000..0d5198a0
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64-crash-reports.tsv
@@ -0,0 +1,2 @@
+crash-reports/vspipe-bin-2026-11-19-095910.ips vspipe-teardown vspipe-bin frameDoneCallback
+crash-reports/vspipe-bin-2026-11-19-101530.ips vspipe-teardown vspipe-bin outputFrame
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64.txt b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64.txt
new file mode 100644
index 00000000..56823ca2
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-macos-x64.txt
@@ -0,0 +1,6 @@
+unpatched trigger (4 loops, CPU load): runs=4000 failures=1
+unpatched ivtc (4 loops, CPU load): runs=3360 failures=1
+unpatched bwdif (4 loops, CPU load): runs=3360 failures=0
+patched trigger (4 loops, CPU load): runs=4000 failures=0
+patched ivtc (4 loops, CPU load): runs=3360 failures=0
+patched bwdif (4 loops, CPU load): runs=3360 failures=0
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-loops-windows-x64.txt b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-windows-x64.txt
new file mode 100644
index 00000000..38949a9c
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-loops-windows-x64.txt
@@ -0,0 +1,6 @@
+unpatched trigger (4 loops, CPU load): runs=4000 failures=1
+unpatched ivtc (4 loops, CPU load): runs=3360 failures=2
+unpatched bwdif (4 loops, CPU load): runs=3360 failures=0
+patched trigger (4 loops, CPU load): runs=4000 failures=0
+patched ivtc (4 loops, CPU load): runs=3360 failures=0
+patched bwdif (4 loops, CPU load): runs=3360 failures=0
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-linux-x64.txt b/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-linux-x64.txt
new file mode 100644
index 00000000..478cf190
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-linux-x64.txt
@@ -0,0 +1,2 @@
+sweep unpatched (fix marker: no): runs=6300 failures=60 (60 exit 1; ); cores: 0 classified (0 the race)
+sweep patched (fix marker: yes): runs=6300 failures=60 (60 exit 1; ); cores: 0 classified (0 the race)
diff --git a/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-macos-x64.txt b/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-macos-x64.txt
new file mode 100644
index 00000000..5eb78a20
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/data/ab-sweep-macos-x64.txt
@@ -0,0 +1,2 @@
+sweep unpatched (fix marker: no): runs=6340 failures=100 (20 exit 139; 80 exit 1; ); crash reports: 20 image:OpenCL,
+sweep patched (fix marker: yes): runs=6340 failures=100 (20 exit 139; 80 exit 1; ); crash reports: 6 image:OpenCL,
diff --git a/docs/reports/task-72-macos-x64-sigill/index.html b/docs/reports/task-72-macos-x64-sigill/index.html
new file mode 100644
index 00000000..84a18040
--- /dev/null
+++ b/docs/reports/task-72-macos-x64-sigill/index.html
@@ -0,0 +1,186 @@
+
+
+
+
+
+macOS x64 sweep SIGILL
+
+
+
+
+Task #72: the macOS x64 format-sweep SIGILL was vspipe's teardown race
+2026-11-19 · GitHub-hosted x86 runners (macos-15-intel, ubuntu-24.04, windows-latest) · deps 1.17.0-rc3 against 1.18.0-rc1 ·
+scripts in scripts/, raw counts in data/
+
+Outcome
+
+- Every macOS x64 SIGILL in the push gate was task #68's race. On deps 1.17.0-rc3, 4 push-gate runs
+failed a random format-sweep row with
vspipe exited with signal 4 (SIGILL) (5 rows). Every one left a crash
+report, and every report has the same faulting thread: vspipe's frameDoneCallback →
+condition_variable::notify_one → pthread_cond_signal →
+_os_unfair_lock_unowned_abort (section 1).
+- It is fixed by deps 1.18.0-rc1, which carries upstream's vspipe fix (task #68). On the same x86
+runners, with only
vspipe-bin swapped, R78's own vspipe crashed 71 of 32,160 one-frame
+runs and the patched one 0 of 32,160 (section 3).
+- No retry or quarantine was added. The sweep's native-crash class still finds real bugs (tasks #69, #70).
+Instead, the CI crash-report step now names this signature and says whether the bundle has the fix (section 4).
+
+
+1. What the crash reports say
+scripts/classify_ci_failures.py reads every failed push-gate job's log (gh run view --job --log),
+lists its failed tests and native-crash rows, and classifies each crash report the log printed with
+print-crash-reports.py's classify(). Output: data/ci-failures.tsv (19 failed jobs,
+deps 1.17.0-rc2 and rc3).
+
+| Platform | Sweep rows that died natively | Signal | Crash reports | Verdict |
+| macOS x64 | 5 (4 runs) | SIGILL | 5 .ips | 5 × vspipe-teardown (frameDoneCallback) |
+| Linux x64 | 10 (9 runs) | SIGSEGV | 1 gdb core (the rest predate core capture) | vspipe-teardown (outputFrame; main in vpy4_freeScript) |
+| macOS arm64 | 0 | | 1 (a Bifrost-era test) | image:AppleParavirtGPUMetalIOGPUFamily, not the race |
+
+The rows were all over the sweep (QTGMC Fast, Crop & Resize, Edge Repair, CTMF, Colour Correction, at different sizes and
+formats), which is what a race in vspipe's exit looks like and what a plugin bug does not.
+
+The three suspected causes
+
+| Hypothesis (task description) | Verdict | Why |
+| A CPU-dispatched instruction the runner lacks | ruled out | No plugin image is on the
+faulting thread; the faulting instruction is ud2 in libsystem_platform, put there deliberately. |
+A Zig ud2 (vszip / zsmooth safety trap) | ruled out | Same: the trap is in
+_os_unfair_lock_unowned_abort, called from vspipe, with neither plugin on the stack. |
+| vspipe R78's teardown race (task #68) | the cause | Every report: vspipe's callback signalling a
+condition variable that main has already freed. |
+
+
+Why the same race is SIGILL on macOS and SIGSEGV on Linux
+R78's frameDoneCallback increments completedFrames before it writes the frame and notifies. Main can
+see "every frame done" in that gap and free VSPipeOutputData, its mutex and condition variable included. What the
+callback then does with freed memory decides the signal:
+
+- macOS:
pthread_cond_signal takes the condition variable's internal os_unfair_lock.
+In freed (or reused) memory its owner field no longer matches, and libplatform aborts on purpose with ud2:
+EXC_BAD_INSTRUCTION, SIGILL. When the callback is still in outputFrame writing to the freed buffer,
+it is a SIGSEGV instead; both A/B crashes on macOS x64 (section 3) were that form.
+- Linux: glibc's condition variable has no owner check, so the fault is the write itself: SIGSEGV in
+
outputFrame, with main in vpy4_freeScript.
+- Windows: an access violation (
0xC0000005).
+
+The fixture Tests/TestResources/crash_reports/macos-x64-vspipe-teardown-sigill.ips is one of these
+reports as CI printed it (paths reduced to the runner's own).
+
+2. Push-gate tally since deps 1.18.0-rc1
+The macOS x64 push gate runs 588 sweep rows a job. On rc3, 5 rows died in 20 jobs (11,760 rows): about 1 in 2,400.
+
+| Run | Branch | macOS x64 |
+| 44411658811 | task #68 | passed |
+| 44481983232 | main | passed |
+| 44484258865 | task #67 | passed |
+| 44497719057 | main | failed: quick-probe 15 min timeout (task #75), no sweep row failed |
+| 44500281558 | task #70 | passed |
+| 44510129664 | task #73 | passed |
+| 44510992864 | task #70 | passed |
+| 44515328673 | task #75 | failed: one 30 s test timeout (task #75), no sweep row failed |
+
+8 jobs, 4,704 rows, no SIGILL. At the rc3 rate 2.0 were expected; the chance of none by luck alone is about 0.14.
+Limit: this is consistent with the fix but would not establish it on its own. The A/B in section 3 is the evidence.
+Corrections: the plan for this task gave P(0) ≈ 0.08 for the first three green runs; at 588 rows a job that is
+1,764 rows and P(0) ≈ 0.47. The plan and the review also counted 5 runs and 6 SIGILL rows on rc3; data/ci-failures.tsv
+has 4 runs and 5 rows (with 5 reports). The eight runs above bring P(0) to 0.14.
+
+3. The x86 A/B
+scripts/repro-vspipe-teardown.yml (the temporary workflow, copied here; run 44500126463). Every job installs the
+deps 1.17.0-rc3 bundle and runs it twice, once with its own vspipe-bin (R78) and once with the
+vspipe-bin from 1.18.0-rc1 (R78 + the patch, checked by its --version marker). Nothing else differs.
+
+Loops
+Task #68's compare_vspipe.sh, 4 loops at once beside CPU hogs: vspipe on trigger.vpy (1,000 runs a
+loop), and the worker's IVTC preview with the pre-#65 template and its Bwdif preview (28 passes over the 30-frame fixture a loop).
+data/ab-loops-*.txt.
+
+| Platform | Script | Runs per side | R78 | R78 + patch |
+| Linux x64 | trigger | 4,000 | 0 | 0 |
+| IVTC | 3,360 | 66 | 0 |
+| Bwdif | 3,360 | 0 | 0 |
+| macOS x64 | trigger | 4,000 | 1 | 0 |
+| IVTC | 3,360 | 1 | 0 |
+| Bwdif | 3,360 | 0 | 0 |
+| Windows x64 | trigger | 4,000 | 1 | 0 |
+| IVTC | 3,360 | 2 | 0 |
+| Bwdif | 3,360 | 0 | 0 |
+| Total | 32,160 | 71 | 0 |
+
+
+- Both macOS crashes left a report; both classify as
vspipe-teardown, one in frameDoneCallback, one in
+outputFrame, both SIGSEGV (data/ab-loops-macos-x64-crash-reports.tsv). The Windows ones were access
+violations (0xC0000005); Windows keeps no crash report on the runner.
+- If the two sides crashed at the same rate, 71 crashes all falling on R78 would have probability 2−71. Without Linux
+(macOS and Windows, 5 against 0) it is 2−5 ≈ 0.03; macOS alone (2 against 0) is 0.25, not significant by itself.
+- The patched rate is below 3 in 32,160 (about 1 in 10,700) at 95% confidence; on macOS x64 alone, below 3 in 10,720.
+
+Limits: the loops are task #65's workloads, not the sweep's. They hit the race much more often on Linux (IVTC: 2%)
+than on macOS (1 in 5,400 over all three scripts), so on macOS they do not reproduce the push gate's rate of about 1 in 2,400, and
+neither macOS crash was the SIGILL form. What they show on macOS is that the race fires there, under the same signature, and that
+the patched side did not crash in 10,720 runs.
+
+Format-sweep replays: inconclusive
+Each sweep job ran the push gate's format sweep once with a capturing vspipe-bin
+(.github/scripts/replay-vspipe.sh), then replayed every captured run 10 times with each vspipe
+(scripts/ab_sweep_replay.sh; data/ab-sweep-*.txt).
+
+| Platform | Captured runs | Runs per side | R78 | R78 + patch |
+| macOS x64 | 634 | 6,340 | 100 | 100 |
+| Linux x64 | 630 | 6,300 | 60 | 60 |
+
+The failures are the same few captured runs failing all 10 times on both sides: KNLMeansCL refusing the runner's OpenCL
+device ("the opencl device does not support this video format", exit 1) and, on macOS, an OpenCL kernel build crashing
+(exit 139, crash reports in image:OpenCL). In the test these rows run with the worker's OpenCL probe, which a
+bare replay skips. No replay hit the race on either side, so this arm says nothing about the fix. Replayed
+alone, without the sweep's other jobs and the test runner around them, a run is apparently too quiet to open the window.
+
+4. What changed
+
+.github/scripts/print-crash-reports.py: prints every thread of a macOS .ips (it printed the
+faulting thread only), classifies a report (vspipe-teardown, image:<name> or
+unknown), and says whether the bundle's vspipe carries the fix (its --version marker read from the
+binary). --classify FILE… runs it on any .ips or gdb text and never fails.
+ci-test.yml and nightly.yml: the macOS jobs upload the raw .ips files as
+crash-reports-* artifacts on failure (14 days).
+app/test/crash_report_classifier_test.dart (push gate): the macOS SIGILL and the Linux SIGSEGV fixtures both
+classify as the race; a plugin/driver crash is named by its image.
+replay-vspipe.sh: runs on macOS (no env -0 or timeout there).
+- CLAUDE.md Debugging Tip 11: the macOS signature.
+
+
+5. Not tested
+
+- The SIGILL form itself under the A/B: the loops on macOS x64 produced only the SIGSEGV form. The SIGILL form's evidence is the
+five CI reports' stacks, which are the same callback.
+- A larger macOS-only A/B. 10,720 runs a side gave 2 against 0; telling the two apart on macOS alone would take roughly
+five times as many runs.
+
+
+Files
+
+scripts/classify_ci_failures.py → data/ci-failures.tsv
+scripts/repro-vspipe-teardown.yml (workflow, removed from .github/workflows/) with
+scripts/ab_sweep_replay.sh and task #68's compare_vspipe.sh → data/ab-*.txt,
+data/ab-loops-macos-x64-crash-reports.tsv
+
+
+
+
diff --git a/.github/workflows/repro-vspipe-teardown.yml b/docs/reports/task-72-macos-x64-sigill/scripts/repro-vspipe-teardown.yml
similarity index 95%
rename from .github/workflows/repro-vspipe-teardown.yml
rename to docs/reports/task-72-macos-x64-sigill/scripts/repro-vspipe-teardown.yml
index 52623d07..b8448873 100644
--- a/.github/workflows/repro-vspipe-teardown.yml
+++ b/docs/reports/task-72-macos-x64-sigill/scripts/repro-vspipe-teardown.yml
@@ -1,9 +1,10 @@
name: Repro vspipe teardown A/B (task 72, temporary)
-# TEMPORARY (task #72): R78's own vspipe against the one with task #68's
-# backported fix, in the same deps bundle (1.17.0-rc3), on x86 runners. Its
-# copy lives in docs/reports/task-72-macos-x64-sigill/scripts/; delete this
-# file before the task's PR is merged.
+# Task #72: R78's own vspipe against the one with task #68's backported fix,
+# in the same deps bundle (1.17.0-rc3), on x86 runners. It ran from
+# .github/workflows/ on the task's branch (run 44500126463) and is kept here
+# with the report; copy it back into .github/workflows/ (and point the push
+# trigger at your branch) to run it again.
#
# loops task #68's compare_vspipe.sh: vspipe on its trigger script, and the
# worker's IVTC (pre-#65 template) and Bwdif previews, 4 loops at