[nmplcpimenta]

Problem

There are two bugs in midi1::input_state_machine::on_bytes_multi_segmented (include/libremidi/detail/midi_stream_decoder.hpp):

1. Wrong copy offset. When a buffer contains other messages before an F0 that is not terminated in that buffer, the partial SysEx is copied with message.assign(bytes.begin(), bytes.begin() + size), which starts at the beginning of the buffer instead of at the F0. For example, 90 3C 7F F0 7E 01 followed by 09 01 F7 gives a note on and then the "SysEx" 90 3C 7F 09 01 F7.
2. Empty buffer. An empty buffer reaches has_finished_sysex, which calls front() and back() on an empty span. This is undefined behaviour, and it trips the libc++ hardening assertion span<T>::front() on empty span. on_bytes_segmented already returns early on an empty buffer.

Change

• The partial SysEx is copied from bytes.begin() + i_byte.
• on_bytes_multi_segmented now returns early on an empty buffer, as on_bytes_segmented already does.

Tests

There is a new case in tests/unit/midi_stream_decoder.cpp, "midi1: SysEx after other messages in one buffer":

• the example above must yield a note on and F0 7E 01 09 01 F7;
• empty buffers passed to on_bytes_multi and on_bytes produce no messages.

On the current master, the first check fails and the second aborts under libc++ hardening.

Relation to the SysEx PR

Related to #271, which rewrites this F0 branch and also never calls front() on an empty span, so it fixes both bugs on its own. The two touch the same lines; whichever lands second, I'll rebase it. If #271 lands first, this PR shrinks to its test, which passes there as is.

🤖 Generated with Claude Code

[jcelerier]

heya! thanks, can you rebase on new master? I merged your other patch + some additional sysex fixes found during review