[LeeO86]

Problem

At 1080p50 with 25 Hz updates, input_grains_total grew at about 47.9/s while the source wrote 50/s. No resyncs were counted, and no other metric showed the missing grains.

Cause

src/mxlio/io.cpp, the input reader loop:

1. mxlFlowReaderGetGrain with a timeout answers MXL_ERR_OUT_OF_RANGE_TOO_EARLY when the wait runs out. It never answers MXL_ERR_TIMEOUT (MXL 218ddaa, PosixDiscreteFlowReader::getGrainImpl with a deadline). The reader thought TOO_EARLY comes back at once, so it paused with wait(5). That pause really sleeps 10 ms, because wait sleeps in 10 ms steps.
2. The wait is cut at the next update tick (timeout = tick - now). So every 25 Hz tick ran into this pause. Then came the update (analyse and render, about 10 ms). Together that often took longer than one frame (20 ms). By then the head was 2 or more grains past the cursor, and the reader jumped to the head (want = head when cursor + 1 < head). The jumped-over grain was not counted anywhere.

Jumping to the head is on purpose: in latest mode an update only needs the newest grain. The pause was the bug. Without it, the reader keeps up unless an update takes longer than a frame.

Change

• No pause after MXL_ERR_OUT_OF_RANGE_TOO_EARLY. It is a wait that ran out, like a timeout, so the loop does not spin. Other unexpected answers still pause.
• New counter mxl_scopes_input_grains_skipped_total for grains the reader jumps over when it is behind. It is counted once, when the grain after the gap is read. accumulate_dropped_total now counts at the same place. Before, a wait that ran out with the cursor still behind could count the same gap again.
• Docs: SPECIFICATION §2.2 and the §10 metrics table, CHANGELOG under Unreleased.

Tests

• Unit tests: 88/88 pass.
• tests/integration/scope.sh has a new check: a 1080p50 scene writer, RGB parade at 25 Hz, 4 s. It wants at least 180 grains read, and read plus skipped between 194 and 206.
  • Before: 175 read, so it fails.
  • After: 202 read, 0 skipped, and the whole script passes.
• Lab, 10 s per run, 1080p50, 25 Hz, grains read per second:

Mode / pattern	Before	After
luma / scene	50.1	50.1
ycbcr-parade / scene	50.1	50.1
rgb-parade / scene	33.4	50.2
rgb-parade / random	–	50.1
rgb-overlay / scene	25.1	50.2

• To check the counter, I built once with the old pause put back. Result: 34.5 read/s + 15.7 skipped/s = 50.2/s.

Note: the new integration check needs one update to take less than a frame on the CI runner (about 10 ms here for the RGB parade). A much slower runner would show skipped grains and fail the 180 bound.

🤖 Generated with Claude Code

https://claude.ai/code/session_017ybuzgBaGxuB5UiaE8dRa4