N6 report: the picture, and the capture path's machine half

Run stamp: 2026-09-06, rustc 1.97.1. Pins: ntsc-crt v0.2.4 by tag (ntsc-grid, ntsc-source-nes, ntsc-decode, ntsc-crt; ntsc-source-cap as a dev-dependency), 2c02 0b25e83 (v2c02-fast by path), 2a03 44277e1, 6502 8b9e0b5, nes-bus v0.1.1. cargo test -p nes-console: 8 tests (N5's six, the two here). Plan: docs/n6-plan.md, written first.

Step 1 is closed and gated. Step 2's machine half runs end to end on this box and is recorded against the tolerances the plan stated: luma holds on every region, hue and saturation do not on 69 of 436, and the cause is located in the capture model, not the console. The real capture of the bars cartridge is the bench item it was.

What the bars cartridge needed first

full_palette.nes paints every palette entry as bars with rendering off: v parked in palette RAM through a $2006 pair, $2007 writes stepping it mid-line, $2001's emphasis bits cycling per band. Through the console it came out as blue stripes, because the fast PPU painted the backdrop with rendering off and said so (a labelled authoring, the P3 worlds never having disabled rendering). Measured on the switch-level 2C02 (2c02: v2c02-dots/examples/blank-probe.rs, rows 60..67 of a frame with mask $00): with v in $3F00..$3FFF the picture is the entry v addresses through the register file's own mirror rule, the backdrop otherwise; a $2006 pair shows five dots after its second write's access starts, a $2007 step eight, the emphasis bit three. The stepper now paints that per dot (blank_colour, the $2007 step held BLANK_2007_HOLD = 3 dots past the write, 2 and 4 red) and carries $2001's emphasis bits beside every dot, held to the capture in 2c02: v2c02-fast/tests/blank.rs (MUTATE=1 backdrop only, red on 1,625 dots). The emphasis lead of two dots is recorded in the 2c02 report and not modelled: the harness puts the byte on the bus at the start of the access, which a 6502's write does not.

Step 1: the picture

nes_console::picture::Picture: each frame encoded by ntsc-crt's NES source at the subcarrier phase the previous frame left, decoded on Rung C (the three-line comb with the NES weights), and through the CRT stages at ntsc-crt's authored parameters, scale 3 (768 x 720). The module adds two things to ntsc-crt's chain, the phase carried across frames and the order; nothing in it is fitted. run-rom writes it with CRT=out.ppm and the bare decoded grid with DECODED=out.ppm. full_palette.nes through it: the bars, the eight emphasis bands, the comb's picture. 30 frames through Rung C and the stages take 2.2 s on one core (13.7 frames a second): the picture is not real time on the CPU, which N8's shell knows from the roof's wasm figures.

Gate (tests/picture.rs, no die data, no goldens):

• A console frame through the picture is the standalone fast rung's frame from the same world through the same picture on every decoded sample: 1,474,560 components equal (2048 x 240 x 3), the two frames at the same parity. The CRT stages run on it and put something on the 768 x 720 screen.
• The phase the picture carries across twelve console frames (four of them short) is what ntsc-grid's arithmetic gives that parity sequence: 4. MUTATE=1 pushes every frame as Even (a full frame's residue is 4, a short frame's 8) and reads 0: red.

Step 2: the capture path, machine half

examples/capture-score.rs: a ROM through the console, its frames encoded in order, and either a real record (the .u8 at the scope's rate the M4 tools read) or the console's last three frames through ntsc-crt's capture-card model at the same rate (125 MHz, 5 ppm off, 20 mV of DC, 2 mV of noise), recovered the way M4 recovers a real one (auto_level_nes, then recover_nes), and scored region by region against the console's own synthesis through the identical decoder (Rung C): mean luma, saturation and hue over the region, one dot in from each edge. Regions are found in the console's frame: the largest flat rectangle of every distinct (colour, emphasis) at least 12 dots wide and 6 rows tall, with a bar's stepped edges merged (full_palette's $2007 timing steps them). On full_palette.nes, 436 regions, 18 of them grey.

Tolerances, as stated in the plan before measuring: luma within 0.01, hue within 1.0 degree where the synthesis has a hue (saturation above 0.02), saturation within 5 percent of the synthesis or 0.005 absolute, whichever is larger; a miss on any region fails the synthetic roundtrip. The real capture is recorded, not held.

What the first run found belonged to the instrument. The luma missed on 130 regions with a bias that grew with the level, a 3.7 percent gain: ntsc-crt's re-referencing placed sync tip and blanking at the centres of 4 mV histogram bins, 1.4 percent of the sync depth at each end. Writing the test that holds it finer exposed a second weakness in the same finder, a picture level below blanking (the NES's row-1 colours dip under it) taken for blanking once it fills a percent of the record, which a solid dark frame does and a bars frame can. Both are fixed in ntsc-crt 0.2.4 (the levels are read as medians inside every sync pulse and on the front porch before it, the histogram only placing the threshold; the M4 report's fourth addendum), the console re-pinned.

With that, the synthetic roundtrip (+0.7 ppm recovered, worst burst residual 0.010):

	regions	within	worst
luma (0.01)	436	436	0.0096
hue (1.0 degree, 418 with a hue)	418	352	12.1 degrees, on a region of saturation 0.02
saturation (5 percent or 0.005)	436	431	0.0078
all three	436	367

69 regions miss, so the synthetic roundtrip does not close at the stated tolerances, and the example exits 1 saying so. The miss is one residual: a chroma vector error of at most 0.0086, deterministic (the same with the noise at zero), the same at the grid's own rate with the rate error at zero, so it is not the resampling or the lock's rate; on the saturated colours it reads as a rotation of +0.4 degrees (the burst lock's residual, 0.010 grid samples, is that angle) and a gain scatter of 1 percent, and on the near-greys the degree tolerance turns 0.005 of chroma into 12 degrees. What is left when the rate and the noise are out is the card model's anti-alias lowpass, 6.5 MHz over 41 taps, on the encoder's square-wave chroma: the synthesis side is decoded unfiltered, and at twelve samples a cycle the square wave's eleventh and thirteenth harmonics alias onto the fundamental the decoder reads, so the two sides hand the identical decoder a fundamental that differs by a percent depending on the colour's waveform. That is ntsc-crt's comparison procedure, not the console's picture, and the plan's rule was not to fit the tolerance to the data: the figures stand as measured, and the procedure question (filter the synthesis with the same front end, or state the tolerance from the aliased harmonics) is named for the next plan.

The real half: no bars record exists (the five records on hand are Super Mario Bros. and Duck Hunt, whose ROMs are not on this box, so no region of theirs can be scored against a console frame). The example takes the record and runs the same procedure; the recording of full_palette.nes on the real console is the bench item, and M4's terminated-capture finding (the probe run flattering chroma by about 40 percent) says the record must be taken terminated.

The procedure question, answered (2026-09-06, the same day)

The plan's rule was not to fit the tolerance to the data, and the report above named a procedure question and a cause. The cause was wrong, and the gate said so as soon as it was tried: the reference side of the comparison now goes through the card model's front end (ntsc_source_cap::front_end, ntsc-crt 0.2.7: the anti-alias lowpass alone, held to the model at the grid rate), which is right in principle (a synthesis compared with a capture must carry the capture's band limit), and it moved nothing: 367 of 436 regions either way. The recovery itself was then held to a front-ended synthesis on a bars frame and reproduced it to 0.0022 V worst and 0.3 percent of chroma gain, so the instrument was not it either. What was left was the scoring's own geometry, and a sweep of the region margin showed it: one dot in from the edge of a sixteen-dot bar is inside the decoder's chroma settling distance, and the filter's transitions were the residual (the worst chroma vector 0.0088 at one dot, 0.003 at two, 0.0001 at four).

Two changes follow, both derived rather than chosen. The margin is now the decoder's: half the decimated lowpass's span plus the two decimated samples the interpolation reaches, five dots (margin_dots, from the Decoder instance's fields). And since blargg's full_palette bars are sixteen dots wide, this repository authored a bars cartridge of its own (testrom::bars_program, exported by export-testrom bars, nobody's game): thirty-two-dot cells, the twelve hues at one luma row and the backdrop, the row stepping every 120 frames through 1, 2, 3, 0, so a run of 100, 220, 340 or 460 frames scores one row.

That cartridge found a third instrument flaw, at luma row 0 only: every hue missed by a rotation that grew with the bar's position on the line, and the recovery read the rate 145 ppm wrong on a 0 ppm capture. The darkest colours' low level, after the card's filter, dips below the sync threshold for half a subcarrier cycle, and every trough was a sync edge to the line walk. ntsc-crt 0.2.8 qualifies an edge by a microsecond of sync after it; the real records are unchanged.

The synthetic roundtrip, on the bars cartridge, with the front end on both sides and the derived margin, at the plan's tolerances (luma 0.01, hue 1.0 degree, saturation 5 percent or 0.005):

luma row	regions	within all three	worst luma	worst hue	worst saturation
1	13	13	0.0001	0.3 degrees	0.0011
2	13	13	0.0001	0.3 degrees	0.0012
3	13	13	0.0001	0.3 degrees	0.0005
0	13	13	0.0001	0.3 degrees	0.0008

It closes. On full_palette the same procedure leaves two regions wide enough to score, both grey, both within: that cartridge cannot carry a hue verdict at this decoder's resolution, which is recorded rather than worked around. The real capture's procedure is the same one, and the bars cartridge is now also the ROM the bench was waiting for.

What stays for the bench

• The bars record: this repository's bars cartridge (export-testrom bars) on the real console, terminated, at the scope's rate, then capture-score bars.nes frames record.u8 125000000 at a frame count inside the luma row on screen.
• The emphasis lead (2c02 report): whether the real $2001 write's emphasis lands two dots ahead of a colour change, or whether the harness's data-at-start access shape made it look so.
• N5's carried items as they stand: the NMI arrival on the scope, gate 3's ROM, the reset hold. (8x16 sprites and the DMC fetch inside sprite DMA closed on 2026-09-06, the second leaving the die's sample address under a collision as a logic-analyser question, 2a03 docs/n3-report.md.)

Pulled at build time from nes/docs/n6-report.md; the repository is the one copy.