Host-side sink stalls starve the USB drain and overflow the firmware histo queue (frame drops on long scans)

Problem

On the 2026-06 long-scan soak the sensor firmware logged ~833 HISTO enqueue fail: queue full events (enq-deq=4, txfail=0). Root cause is host-side: the ScanRunner consume loop runs on the same thread that drains the live USB histogram batch queue, so a slow disk/DB sink (SQLite commit/fsync, CSV flush) stalls the drain → backpressures the parser → dev.read → the firmware's deliberately shallow 4-deep (~100 ms) histo queue overflows and frames are dropped. Brief (>100 ms, <1 s) host stalls — not a USB fault, not the persistent histo-wedge.

Fix

• AsyncSink: run storage sinks' consume() on a dedicated worker thread behind a bounded buffer, so transient sink stalls never touch the USB path (sustained overload still backpressures — correct).
• Deepen LiveUsbSource._batch_queue to an explicit ~3 s budget computed at the true per-camera row rate (CAPTURE_HZ × active_cameras).
• Instrument StreamInterface packet-queue puts so sub-second host stalls and drops are counted, warned, and summarised per scan (previously invisible below 1 s).
• Document the sink consume() latency budget and shared-payload contract in docs/API.md / the Sink protocol.

Out of scope: bumping firmware HISTO_QUEUE_SIZE (RAM-bound on the sensor; host-side cause removed instead).

PRs

• #96 — original implementation (superseded; close when #113 merges)
• #113 — revised: fixes the ~16× batch-queue sizing error, the AsyncSink join-timeout finalize race (could delete a DB session as "empty"), stale-queue reuse poisoning, and warn-token cross-suppression; rebased onto next through #111/#108

Verification

• tests/test_pipeline/, tests/test_scan_workflow.py, tests/test_stream_interface_backpressure.py: 274 passed / 0 failed on the rebased branch.
• HIL follow-up: repeat a long-scan soak and confirm zero HISTO enqueue fail events and a clean backpressure summary in the stop_streaming log line.

🤖 Generated with Claude Code

Start by reviewing revised PR #113 and the AsyncSink, LiveUsbSource._batch_queue, and StreamInterface entry points named in the issue. Run the listed pipeline, scan-workflow, and backpressure tests, then repeat the long-scan HIL soak; done means the tests pass, no HISTO enqueue failures occur, and stop_streaming reports clean backpressure.