Core pacing & estimator fixes: chunk-duration recompute, estimator drift, write quantum, deadline sleeps

[SpaceK33z]

Fixes the shared-core findings from the DAC output-quality analysis (docs/dac-quality-analysis.md on claude/dac-output-quality-cf26w4). All behavioral fixes come with regression tests; cargo fmt/clippy clean, full test suite green (438 lib + 11 doc tests, 9 new).

Output quality

• UdpTimed: recompute the send period whenever pps changes, not only when the chunk size changes. Previously, in the clamped regime (chunk pinned at max_points_per_chunk, e.g. IDN's 179 points at pps ≥ 17.9k) a pps change kept a stale send period → feeding at the old rate → continuous underrun.
• UdpTimed: bounded stall catch-up — after a scheduling stall, next_send may lag now by up to 3 chunk durations so lost time is worked off, instead of being discarded by the old next_send = now clamp (which permanently drained the device buffer on every hiccup).
• SoftwareDecayEstimator: carry fractional consumption across rebases. record_send used to floor-truncate ~0.5 points per rebase (one-sided), inflating the estimate by ~500 phantom points/sec at 1 kHz write cadence and pinning the real device buffer at zero. New multi-rebase drift test asserts <1.5 pt drift over 1000 rebases.
• NetworkFifo: minimum write quantum (max of 5 ms of points / 16, capped to the target buffer) — stops the ~1000 tiny-writes/sec steady state that amplified estimator drift and per-write overhead.
• Absolute-deadline sleeps in sleep_with_control_check and the reconnect backoff — sleep loops no longer assume thread::sleep returns on time (it can overshoot 10x+ on Windows' default timer).
• Frame-capacity clamp for authored frames: frames whose authored points alone exceed frame_capacity (Helios: 4095) are now truncated with a rate-limited warning instead of being sent oversized to the device.
• IdlePolicy::Park honored in the armed-no-frame path (previously always blanked at origin).

Reliability / safety

• Control messages processed inside the NetworkFifo WouldBlock spin — a Disarm (shutter close) previously waited indefinitely while a backend stalled. Also adds a 2 s staleness bound: continuous WouldBlock now escalates to disconnect so reconnect engages instead of spinning forever.
• Reconnect attempts first, then backs off — a transiently recoverable disconnect no longer blanks output for a full backoff period before attempt feat!: streaming-first API refactor #1.
• Scheduler thread is now named (laser-frame-scheduler); priority elevation left as documented future work (no new dependency).

Docs

• CONTEXT.md OutputModel examples corrected (LaserCube WiFi and AVB are NetworkFifo, UdpTimed's example is IDN) and the discard_cached contract updated to match reality (on_reconnect owns cache disposal).

Notes for review:

• Tunable constants introduced: MIN_WRITE_QUANTUM_SECS=0.005, MIN_WRITE_QUANTUM_FLOOR=16, WOULDBLOCK_STALL_TIMEOUT=2s (network_fifo.rs), CATCH_UP_CHUNKS=3 (udp_timed.rs).
• The write quantum is capped to target_buffer_points so very low pps sessions (tiny target buffers) can't wedge — this was caught by an existing pps=30 metrics test.
• Skipped (low severity, deliberately): rotational color delay for cyclic frame-swap slices — invasive relative to its impact.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N4AZhstwBFe428pC7cJMbT

---

Generated by Claude Code