capnproto: bump to 1.4.0

[greatgitsby]

Matches the capnproto bundled by pycapnp 2.2.4. Context: commaai/openpilot#38761

After this merges and the wheel is published: bump pycapnp to 2.2.4 in openpilot (drops the 2.1.0 pin) and re-lock comma-deps-capnproto.

Verified locally: capnproto/build.sh, installed into an openpilot venv, regenerated capnp headers, full tools/op.sh build, cabana demo route 30s, tests/test_cabana.

pycapnp 2.2 memory leak

openpilot pinned pycapnp==2.1.0 (commaai/openpilot#38126) because 2.2 leaked in realtime processes (GC disabled). Root cause: from_dict touched self.schema.fields for every kwarg, and that wrapper forms a reference cycle with the builder, so each message is kept alive until a GC pass. Fixed upstream in capnproto/pycapnp@a27c849 (#401, released in 2.2.3) by only doing the lookup for str values.

Repro, GC disabled, 300k messages per case, fresh process per case:

import gc, resource
from openpilot.cereal import log
gc.disable()
for i in range(300000):
  log.Event.new_message(valid=False, logMonoTime=i)   # int/bool kwargs
  # log.Event.new_message(logMessage='x')             # str kwarg
print(resource.getrusage(resource.RUSAGE_SELF).ru_maxrss // 1024, "MB", gc.collect())

pycapnp	bundled capnproto	int/bool kwargs	str kwarg
2.1.0	1.0.1	0 MB	0 MB
2.2.0	1.0.1	+9056 MB	+8816 MB
2.2.1	1.0.1	+9057 MB	+8816 MB
2.2.2	1.0.1	+9056 MB	+8816 MB
2.2.3	1.0.1	0 MB	+8816 MB
2.2.4	1.4.0	0 MB	+8817 MB

The str kwarg path still leaks in 2.2.4 (the cycle itself was not fixed, only sidestepped). In openpilot the only caller passing a str kwarg is logmessaged (logMessage=record), which does not run with GC disabled. Every config_realtime_process user goes through messaging.new_message, which only passes valid and logMonoTime.

[adeebshihadeh]

Don't we just have to bump pycapnp in openpilot's uv.lock?

[greatgitsby]

i suppose so, but then we'll be using different versions of capnp between py and cpp.

separately, bumping pycapnp won't cause a mem leak like before, but there is still mem leak risk:

The str kwarg path still leaks in 2.2.4 (the cycle itself was not fixed, only sidestepped). In openpilot the only caller passing a str kwarg is logmessaged (logMessage=record), which does not run with GC disabled. Every config_realtime_process user goes through messaging.new_message, which only passes valid and logMonoTime.

capnproto/pycapnp#407 will fix it upstream