6502tinymachines

P0 report: the netlist loads and settles

Run stamp: 2026-09-02, first commit of this repository, rustc 1.97.1, halfphi v0.1.2 (the published crate, untouched), cargo test --workspace 3 tests green, MUTATE=1 red where it must be. The handoff sketch is docs/ppu-handoff-v0_1.md, ratified by the director 2026-09-02 with the repository named tinymachines/2c02.

What closed

  • The 2C02 is the fourth chip through halfphi's identical calls, with zero parser changes: parse(ChipSource { rails: gnd/pwr }), Netlist::decode, Engine::new. The third rail spelling in the family (6502 vss/vcc, 6800 gnd/vcc, 2C02 gnd/pwr) is why rails were a parameter.
  • Counts, agreed by both real parsers: 16,758 transistors, 8,770 nodes (ids to 10,905), 858 names. The sketch's first number, 16,871, was a regex counting the 114 transistors Quietust left inside a block comment in transdefs.js; a second comment-aware regex said 16,757, fumbling one entry. halfphi and the reference JS engine itself agree on 16,758, and the counts test records the saga: instruments lie before the code does.
  • Power-on converges with no chip-specific help (zero nonconvergent settles), the same claim the 6800 and Z80 make.
  • One piece of real chip knowledge was needed, and it is the finding of the milestone: the 2C02 has 38 supply-gated transistors (the 6502 has none). In silicon they conduct permanently; the reference's init turns them on; halfphi's power-on starts every transistor off, and since rails are never recalculated nothing would ever switch them. Ppu::power_on sets exactly those 38 conducting once, documented in place, and the MUTATE=1 proof switches them back off: the replay then diverges at step 0 (an earlier mutation floated the io_ce pin instead, and the node simply held its charge; a mutation the subject survives by design proves nothing, so it was replaced).
  • The reference trace replays. tools/golden-trace/gen.js runs Quietust's own engine headlessly (wires.js + chipsim.js only, the init and stepping implemented statement for statement from macros.js, both buses undriven on both sides), 601 states over 10,906 node slots. The Rust engine matches bit for bit on 10,897 of 10,906 nodes across all 601 states.

The nine that differ, and why the list is closed

Nine sprite-path input latches are dynamic storage with no reset connection: their power-on state is genuinely undefined, and the two engines flip that coin deterministically but differently. Measured: six (the sprite data-in latches and their derived terms, two of them unnamed) flush to agreement by half-step 14, when real values first move through them; three (the x-flip input latch and its followers) are never written in a memory-less free-run and hold their coin for the whole trace. The golden test exempts exactly these, by name where names exist, with the early/late split at the measured flush step, and any tenth node fails the run: a masked comparison is only as honest as its mask is small, and the 6502's rail-write bug hid behind exactly this shape of blindness.

Throughput, measured (the sketch's estimate replaced)

Rung 0 runs at 50,457 half-steps/s on the family's Ryzen 5 5600X: faster than the 6502's 29,600 despite 4.8x the transistors, because a reset chip with rendering off is mostly quiescent and a settle walks only the active cone. A full frame at this state is about 14 seconds, not the sketch's authored "minutes". Rendering on will be slower and P1 will measure it rather than guess.

Carried forward

  • P1: the bus harness (VRAM, palette, OAM, registers), free-running render, the DotFrame extraction, first light through ntsc-crt, and the DAC tap (vid_luma*, vid_burst_*, vid_emph are all named nodes) held against data/nes-levels.toml with tolerances stated first.
  • The x-flip latch trio will leave the exemption list the moment P1 writes real sprite data through it; the exemption shrinking is itself a P1 assertion worth making.
  • Open from the sketch: whether DotFrame moves to a shared contract crate, and the Quietust licence courtesy note.

Superseded 2026-09-04: there was no undefined latch

The nine-node exemption above was the engine, not the silicon. halfphi 0.1.6 resolves a group with no rail and no pull by the reference's own area vote (until then any one charged member made it high; the 2C02's getNodeValue weighs the members' areas), and under it the P0 golden replays bit-exact on all 10,906 nodes across all 601 states with no mask, the x-flip trio included. P1's wider family closed the same way (docs/p1-report.md, its section of this date); the write-path divergence that led there is in docs/p3-report.md. The test compares every node; MUTATE=1 still goes red at step 0.

Pulled at build time from 2c02/docs/p0-report.md; the repository is the one copy.