[facontidavide]

Why now

These three are decoder and schema decisions that 1.0.0 would freeze. Fixing them later changes how already-valid bytes are read.

1. Unknown fields inside a PointField

decodePointField read a varint before it knew the field number, then returned "unknown", and parseFields skipped a second field. Adding any varint field to PJ.PointField in a later release would have broken PointCloud, GridMap and VoxelGrid decoding on every 1.0.0 host.

parseFields now applies one rule to every codec:

Handler returns	Meaning	Loop does
true	field consumed	continue
false, nothing read	unknown field or unexpected wire type	skip it
false, something read	known field is malformed	fail the decode

Before, the third case skipped the bytes that followed and could decode as something else. I audited the handlers of all 20 decoders for a read that precedes the field check; decodePointField was the only one.

2. Absent fields are the proto3 zero value

A standard protobuf encoder omits a zero. Decoders that started from the C++ struct defaults read:

Sent	Decoded before
PointCloud.is_dense = false	true
PointCloud.height = 0	1
PointField.count = 0	1
CylinderPrimitive.top_scale = 0 (a cone)	1 (a cylinder)
annotation colour with a = 0	opaque

They now start from the zero values, and the empty PointCloud payload agrees with them. ImageAnnotations colours use the shared colour decoder. A NaN colour channel decodes as 0 instead of reaching an undefined float-to-integer cast.

Bytes written by the SDK's own serializers are unaffected: they write every one of these fields.

Also in this class, found by a Codex review of this PR:

• An absent annotation thickness, circle diameter or font_size decoded as 2, 1 and 14. They now decode as 0. A thickness of 0 is a real value (a filled shape with no outline) that a standard encoder omits. A consumer that wants a default for 0 applies it when drawing.
• deserializeBuiltinObject(kMeshMap, ...) accepted the empty payload as a default mesh with position_type FLOAT32, while any other payload that omits the field is rejected. The empty payload is now rejected too.

One exception remains, on purpose: a colour (or Mesh3D.scale) that is absent as a whole submessage keeps its documented default.

3. Log.category moves from field 5 to 7

Fields 5 and 6 are file and line in foxglove.Log, so a Foxglove-encoded log decoded its file path as the category. Both are now reserved. Only 1.0.0-dev.1 shipped field 5; a log encoded by a dev.1 build loses its category.

Verification

• The first commit's nine tests were run against the unchanged decoders first: eight failed there. The ninth (NaN channel) passed, because the old behaviour was undefined rather than visibly wrong on this compiler.
• The second commit's tests were not separately run against the old code.
• With the fixes: ./build.sh --debug && ./test.sh gives 107/107 under ASAN. Pre-commit is clean.
• A Codex review read all 43 parseFields call sites and the hand-written loops and found no handler the new rule misclassifies. Its two findings are the second commit.
• Not run: ./test_sdk_install.sh (no packaging change).

docs/builtin_type.md now states the two decode rules.

🤖 Generated with Claude Code