6502tinymachines

A0 report: the fifth chip loads, settles, and replays with no exemptions

Run stamp: 2026-09-02, rustc 1.97.1, halfphi v0.1.3 (released for this milestone; see the finding below). cargo test --workspace --release with REQUIRE_NETLIST=1 REQUIRE_GOLDEN=1: 3 tests green (counts, convergence, the golden). MUTATE=1 goes red: the supply-gated fix-up switched off diverges from the golden at step 0.

What closed

  • The netlist: Quietust's Visual 2A03, eight files fetched by pinned sha256 (recorded at first fetch, 2026-09-02). Both real parsers agree: 10,946 conducting transistors over 5,577 defined nodes in a 20,001-id space, 2,021 names, rails spelled gnd/pwr like the 2C02.
  • Power-on converges unaided, zero nonconvergent settles.
  • The node golden replays bit-exact on every node across 601 states with no exemption list at all, the first chip in the family to replay with none (the 6502 carries the rail encoding caveat, the 2C02 its nine reset-less latches). The unit is the MASTER half-step, one clk_in toggle; the ÷12 clk0 is an output the divider produces, and the golden generator restates macros.js initChip statement for statement with the buses undriven.

The finding: halfphi's drive order, finally exercised

The first replay diverged on exactly eight nodes, every one inverted for all 601 states, two of them named: so and c_so (examples/a0-diverge-probe.rs). The cause is the case the family had documented and never seen: Stats::contested_groups, zero on every chip since the count was built, is 3 on the 2A03 from power-on. Quietust's init drives the so node low while its group carries a layout pullup; the reference resolves such a group by first match walking out from the driven seed (low), and halfphi's old PullDown < PullUp order resolved it by maximum (high).

halfphi 0.1.3 swaps the two: PullDown outranks PullUp, the external drive beating the depletion load, which is both the physical reading and the reference's observed one. With the swap the replay is bit-exact, zero divergent nodes. The swap is proven unobservable everywhere else: contested is 0 on the 6502, 6800 and Z80 (asserted in halfphi's chips test from 0.1.3 on) and the 6502 workspace's full 117 tests, goldens required, re-proved green before the release was cut (digest d1c6947ca960e9eb, tags v0.1.3 there and halfphi-v0.1.3 in the 6502 repo). The 2C02's suite re-proves green on the 0.1.3 pin in its own repository.

  • The supply-gated family recurs, bigger: 100 transistors gated by the supply rail (2C02: 38, Z80: 32), permanently conducting in silicon, set so by power_on and proven load-bearing by the MUTATE red above. The 67 ground-gated are permanently off in both models.

Throughput

114,740 master half-steps/s quiescent, best of three (examples/bench.rs), buses undriven. Faster than the 2C02's 50,457 because a master half-step moves less of this chip: the ÷12 keeps most of the core still per tick. Real time is 42.95 M master half-steps/s, so rung 0 is 374x slow, which is what the ladder is for.

Carried forward

  • Promotion of the supply-gated fix-up into halfphi is deferred, deliberately. The sketch said "if it recurs, promote"; it recurs, but folding it into halfphi changes the Z80's recorded baseline (its 32 supply-gated transistors are documented off there, and its convergence was measured with them off), so the promotion wants its own measured change across all five chips, not a rider on 0.1.3.
  • A1/A2 (the die pages) come after the ladder, per the console sketch v0.2's decision record; the probes they would have hosted are N3 deliverables.
  • A3 (first sound: the AD1/AD2 taps against the wiki's mixer table) is next in this repository.
  • The harness (memory, the 24-edge-style register protocol, DMA) and a golden through it arrive with the pin-lockstep gate against v6502-pins rung 3, the console sketch's "chip versus chip through the contract".

Pulled at build time from 2a03/docs/a0-report.md; the repository is the one copy.