[kahrendt]

Summary

Two rare faults from #24: a segment's boundary time overflowed after about 106.7 days of one anchor, and a live source lost up to a chunk of audio after its ASRC refused one.

Change

• sched::Segment holds its time from T0 as elapsed_s plus elapsed_ps (the picoseconds past the whole seconds, below 10^12) instead of one int64 count of picoseconds. add_segment() carries the remainder into the seconds; segment_output_index() rounds the absolute boundary instant as before, so rounding still never accumulates across segments and picosecond precision stays.
• The 64-bit math still runs only in add_segment() (per set_format() and anchor build) and segment_output_index() (per instance start and format boundary): no multiply or divide is added per sample, cell or frame, and segment_output_index() loses two divides.
• LiveSourceSink::correct(): when the ASRC refuses a chunk, the chunk is written unaltered and the correction restarts (engaged or not), instead of advancing written_ past an ASRC position that did not move, which cut the next output.
• SinkRig moves to tests/source_sink_rig.h, shared by the sink's two test files. Folded: SourceAsrc::produced()'s doc now mentions reset().
• Left for the owner: the same kind of overflow in offset_us() and time_position() within one segment after about 278 days at 384 kHz, which runs per frame: Keep a segment's time and position math from overflowing in a run of months #94.

Tests

• SchedTimeline.SegmentOutputIndexMonthsIntoARun: a Control row at one hour, rows just under and past the old limit (106 and 107 days), a floored remainder, half-up rounding of the remainder and a carry into a whole second, asserting the split fields and the output index. Before the change the 107-day rows trip UBSan's signed overflow.
• LiveSourceSink.StartsTheCorrectionOverAfterTheAsrcRefuses and ...BeforeIt (new test_source_sink_internals.cpp, white-box): a chunk refused once the correction has started, and one refused before it; both fail before the change (about a chunk lost). The refusal is forced by a zero block size, since no real input makes micro-resampler refuse.
• Known gap: the ASRC tests skip in the embedded profile, which corrects by slipping frames.

Closes #24
Related: #94

[Copilot]

🟢 Approval recommended

The focused fixes address both reported faults with targeted regression coverage and no identified blocking issues.

0 open findings

What changed in this PR

Addresses #24 by preventing long-running segment boundary overflow and restarting live-source correction after a resampler refusal. The separate within-segment overflow remains tracked in #94.

Changes:

• Splits boundary elapsed time into seconds and picoseconds while preserving rounding.
• Restarts correction after writing a refused chunk unaltered.
• Adds regression tests and extracts a shared sink test fixture.

File	Description
tests/​test_source_sink.cpp	Uses the shared sink fixture.
tests/​test_source_sink_internals.cpp	Tests refusal recovery before and after correction starts.
tests/​test_sched_timeline.cpp	Tests long-duration boundaries, rounding and carry.
tests/​source_sink_rig.h	Provides the shared sink fixture and tone helpers.
tests/​CMakeLists.txt	Registers white-box refusal tests.
src/​source_sink.h	Documents correction restarts.
src/​source_sink.cpp	Resets correction after resampler refusal.
src/​source_clock.h	Clarifies the produced-frame counter documentation.
src/​sched/​timeline.h	Stores split boundary elapsed time.
src/​sched/​timeline.cpp	Accumulates and rounds split boundary time.
CLAUDE.md	Documents the shared test fixture.

🧠 Review effort: Balanced

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.