[jackulau]

Measured first, rather than assuming

USB traffic at 121 KiB/s SPI : 246 KiB/s = 2.01 Mbit/s
USB full-speed ceiling       : 12 Mbit/s  ->  we use 17%
SPI bits/s at 121 KiB/s      : 0.99 Mbit/s

USB is not the limit. The ~1 MHz SPI clock is.

The headroom I am deliberately NOT taking

We never configure the CH341A's stream clock. Init sends only CMD_UIO_STREAM pin direction; there is no stream-speed command anywhere in the backend, so we run at the device's power-on default. That is genuine unexploited headroom.

It is left alone on purpose:

1. The value is unmeasured, and this project has a hard-won lesson about theorising from reference source instead of benching (a CS-deasserted bench needs no chip contact and is free and safe, so it should come first).
2. A faster SPI clock degrades signal integrity on exactly the marginal probe this whole effort exists for. Speed and contact are in direct tension here. Shipping a clock change blind could make reads worse while looking like an optimisation.

Worth doing, with a bench, on real silicon. Not worth guessing.

The number that actually matters

Throughput is not the binding constraint. A chunk only counts if it finishes inside ONE unbroken contact window. At 121 KiB/s, with backup chunks read twice:

chunk	contact needed
64KB	1.058s
32KB	0.529s
16KB	0.264s
8KB	0.132s
4KB	0.066s

A probe that cannot hold for 1.06s makes zero progress at 64KB, no matter how many windows it gets: every chunk dies partway and is discarded. That is a cliff, not a slope, and no clock speed climbs it. Measured contact on the board is 1 sample in 32.

So: ratchet read --chunk-kb N, default 64 (no behaviour change unless asked for).

Changing it restarts the dump. The sidecar records the stride and refuses to mix two griddings of one chip; that rejection is pre-existing paranoia and it is right here, because bytes banked at one stride were never verified at another.

The test was weak and mutation testing caught it

The first version asserted against a ResumeState the test itself constructed. It proved nothing about the read, and stayed green when set_read_chunk was mutated to ignore its argument: a decorative test for a decorative knob.

Rewritten to observe the grid the read wrote, via the sidecar it saves on failure. Both mutations now fail it:

Mutation	Result
set_read_chunk ignores its argument	FAILS (was: passed)
read loop uses the const instead of the field	FAILS

CLI bounds verified end to end: --chunk-kb 99999 and --chunk-kb 0 both rejected.

Verification

cargo test --workspace 501 pass, clippy --workspace --all-targets -D warnings clean, fmt --check clean.