[jcelerier]

Draft: depends on ossia/libossia#957. The submodule here points at that PR's branch commit (74fa1a3); it must be re-pointed at libossia master once #957 lands, and this PR cannot go green before then.

The bug

Every number a QML script puts in an OSC message goes on the wire as a float, so there is no way to send the int source index SpatGRIS's /spat/serv grammar requires. ossia::qt::value_from_js has a lone numeric branch and QJSValue::toNumber() is a double. The shipped Spatial Protocol Mapper, which builds that message in QML, emits (oscdump, released package):

/spat/serv sffffff "deg" 1.000000 -90.000000 0.000000 1.000000 0.400000 0.600000

score's own C++ SpatGRIS protocol gets this right — SpatGrisSourceOffsetTest.cpp asserts IsInt32() on that argument — so the QML path is the odd one out.

libossia#957 adds the opt-in third argument osc(address, values, typetags), the type tag string of liblo's oscsend. It is opt-in rather than inferred from the value because a JavaScript number carries no integer-ness: inferring would make the shape of the packet depend on the data (siiiiff for a source on an axis, sifffff a moment later), which is worse than being consistently wrong. The reasoning is in that PR.

This PR

The score-side regression test, because this is where the Mapper + loopback-socket harness lives.

• tests/testdata/protocols-qml/test_osc_typetags.qml — a Mapper that sends the /spat/serv shape with "sifffff", one argument per supported tag ("ifsTF"), a message with no tag string at all, and a tag string shorter than the argument list.
• MapperDatagramProtocolsTest.cpp — asserts the exact datagram bytes against hand-written OSC fixtures (address and type tag strings padded to four bytes, network-order IEEE-754 payloads), so nothing libossia encodes is its own oracle, and reads the type tag string back out of the packet with its own four-byte-padding arithmetic.

The untyped case is there to pin the default: it must stay ,f byte for byte, or every script and stored document already out there changes meaning.

Failing without the fix

With the submodule back at libossia master, QML resolves osc(address, values, typetags) to the two-argument overload and silently drops the third argument — i.e. exactly the shipped behaviour:

MapperDatagramProtocolsTest.cpp:343: FAILED:
  CHECK( oscTypeTags(spat) == ",sifffff" )
with expansion:
  { ',', 's', 'f', 'f', 'f', 'f', 'f', 'f' }  ==  ",sifffff"

... index byte-compared as 3f800000 (= 1.0f) where 00000001 was wanted

test cases:  26 |  25 passed | 1 failed
assertions: 425 | 419 passed | 6 failed

Exactly the 6 typed assertions fail; the untyped ones pass in both builds, and so do the other 25 cases in the binary.

Wire evidence

oscdump — an independent liblo decoder — reading the datagrams this Mapper actually emitted:

/test/all      ifsTF    1 2.500000 "hi" #T #F
/test/partial  if       3 4.000000
/spat/serv     sifffff  "deg" 1 -90.000000 0.000000 1.000000 0.400000 0.600000
/test/untyped  f        7.000000

🤖 Generated with Claude Code

https://claude.ai/code/session_019QYyvhRXvSMSRQPaq12Teq

[jcelerier]

Closing in favour of #2323 for the consolidated engine branch.

The branch fix/qml-osc-typetags is kept, not deleted: it holds the ready-to-land score half of S5 (commit a76778e — the test_osc_typetags.qml fixture, the byte-exact wire assertions in MapperDatagramProtocolsTest.cpp, and the 3rdparty/libossia gitlink).

It is deliberately not on fix/engine-2026-10 yet: the test cannot pass without ossia/libossia#957, and putting an unmerged submodule sha on the shared branch would break every other checkout. Once #957 merges, re-point the gitlink at libossia master and cherry-pick a76778e onto fix/engine-2026-10.

[jcelerier]

Recovered: the test from this PR is now in #2330 as 3348a8fd54, with the submodule pinned to d94955801 on libossia master (the commit ossia/libossia#957 merged as) instead of to the pull request branch — which is what made this one unmergeable.