cdmc: correct the IndyCam register map and power-on defaults

[hanshuebner]

Live camera capture ([vino] source = "camera") came out uniformly green — luma and geometry correct, red and blue crushed to near zero. Reproduced on IRIX 6.5.22 with vidtomem -f /tmp/cap -v 0.

Two compounding bugs in the emulated CDMC:

The register map was wrong on every subaddress except VERSION (0x0E). The IndyCam map — same silicon the Linux indycam driver drives — is:

Sub	Register	Access	Default
0x00	CONTROL (AGCENA, AWBCTL, EVNFLD)	rw	AGCENA
0x01	SHUTTER	rw	0xFF
0x02	GAIN	rw	0x80
0x03	BRIGHTNESS	r	0x80
0x04	RED_BALANCE	rw	0x18
0x05	BLUE_BALANCE	rw	0xA4
0x06	RED_SATURATION	rw	0x80
0x07	BLUE_SATURATION	rw	0xC0
0x08	GAMMA	rw	0x80
0x0E	VERSION	r	0x10
0x0F	RESET	w	—

Mapping VERSION to 0x00 also meant every CONTROL write was dropped as a write to a read-only register.

The register file powered up at all-zero while apply_uyvy_field treated 0x80 as each control's neutral point, so every unprogrammed control sat at its extreme. The balance and saturation terms pulled Cb and Cr from 128 down to 96, which through BT.601 is R−51, G+38, B−64 per pixel — exactly the observed cast. Note the hardware defaults are not uniformly 0x80, so "0x80 is neutral" is wrong even with a correct map.

Fix

apply_uyvy_field is anchored at the power-on values: with the defaults in place every term is unity and the field passes through byte-identical. Controls respond proportionally as the guest moves them away from default. AGCENA and AWBCTL suppress the gain and balance terms respectively, matching a host camera running its own auto-exposure and auto-white-balance. RESET and read-only enforcement are handled.

Verification

Isolated host capture from the guest first: a probe through the same nokhwa path showed macOS AVFoundation delivering correct 1920×1080 packed YUYV, clearing src/camera.rs. Replaying that real frame through Cdmc::apply_uyvy_field offline with candidate register states reproduced the reported green image exactly from the "gain set, everything else zero" state, and produced a clean frame with correct defaults.

Six new tests in src/cdmc.rs, including power_on_defaults_pass_the_field_through_untouched. Full suite passes (308).

Write-up in rules/irix/indycam-cdmc-register-map.md.

Not addressed

The live preview in the video tool is still grey. That is the separate continuous-capture gap noted in rules/irix/indycam-end-to-end-capture.md — read_channel_reg's field_counter >= 2 descriptor-table readback and dma_emit_dword's field_counter == 0 defer are one-shots tuned for a single interlaced frame, so nothing is delivered once a channel streams past the second field.

Also unconfirmed: which subaddresses the IRIX driver actually writes. The cast is cured either way since the defaults are now pass-through, but confirming the map end-to-end wants a run with log vino on.

🤖 Generated with Claude Code

[hanshuebner]