F3 derived the drain_inference timeout from the CHUNK_SAMPLES constant
and capped it near 6s. That breaks on the realistic worst-case
configuration (slow CPU + Whisper large-v3, 3-5x realtime): a 4-second
chunk legitimately takes ~20s to clear the decoder, but the F3 budget
aborted it after 6s and treated healthy work as a wedge — the lifecycle
keeps surviving, but every long-tail chunk gets cancelled and the user
sees their final stretch of dictation get dropped.
The deadline now derives from the in-flight task's own duration_secs
multiplied by a REALTIME_SAFETY_MULTIPLIER constant (5x — the
documented upper bound for the slowest supported backend), with a
DRAIN_TIMEOUT_FLOOR of 2s so sub-second tail chunks still get enough
wall-clock to amortise model load, OS scheduling jitter, and the
abort-callback's own poll cadence. Both constants sit next to
CHUNK_SAMPLES with doc comments explaining the rationale.
Defensive: non-finite or non-positive durations fall back to the floor
so a malformed task can't produce a NaN/overflow budget.
Regression tests (commands::live::tests):
- drain_timeout_scales_with_inflight_chunk_duration_secs: 4.0s chunk
must get 20s budget (5x), not the old 6s cap.
- drain_timeout_honours_floor_for_short_chunks: 0.3s chunk produces
1.5s scaled value, must be clamped to the 2s floor.
- drain_timeout_uses_floor_when_no_inflight_task: well-defined fallback
for the (in practice unreachable) None branch.
- drain_timeout_rejects_non_finite_duration: NaN / inf / 0 / negative
fall back to floor.
cargo test -p lumotia --lib commands::live::tests::drain: 4 passed.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Jake
07f6755961