feat(matrix): keep everything a run measured, not only its headline

[drewstone]

A row answered "how did it do". Every autopsy in this work has needed "what happened" — and the data was already collected, then discarded at the row boundary. latencyMs was a full per-decision array. The evidence watcher held every channel at every decision. Driver health was a structured record being flattened into one sentence of prose.

The raw trace

Every decision, in order: the turn, the word played, that decision's latency, and every evidence channel as it stood afterwards.

turn 1 input=NOOP latency=82ms score=0 lives=5 ball=(0,0) paddle=146
turn 2 input=FIRE latency=0ms score=0 lives=5 ball=(124,116) paddle=146

On ALE Breakout that includes the ball and paddle RAM addresses, so you can watch the ball spawn at the exact decision that pressed FIRE. Written beside the study as JSONL, one decision per line — a 20,000-decision episode inside the row would make the study file unopenable.

Distributions, not means

spreadOf reports n, min, p50, p90, p99, max, mean and a twelve-bucket histogram.

Measured on one cell: 125 decisions at ~0 ms and one at 329 ms, for a mean of 3.20 that describes neither. A mean has hidden every interesting result in this work, including a profile that looked consistent to 0.2% at a censored horizon and varied 5.3× uncensored.

Also on the row now

The scored channel as a series; every channel's last value; input total against distinct inputs; per-milestone verification; the transport's health as data rather than prose; budget exhaustion; and the build phase's input tokens, output tokens and turns kept apart. That ratio is diagnostic — 1.8M in against 72 out is an agent reading, not one writing.

Artifacts survive

The authoring sandbox is a temp directory that gets reclaimed, so the program an agent wrote and the transcript of it writing were lost when the run ended. Both are copied beside the study now.

It found a defect immediately

The bundled masher emitted a hardcoded lowercase word. On Breakout, whose vocabulary is NOOP FIRE RIGHT LEFT, every input was illegal and the first decision spent ten seconds reaching a driver timeout. It reads its vocabulary from the request now, as the cycler already did — I had fixed one of the two controls and the trace caught the other.

Local: 20 of 20 test files, boundary check, 0 typecheck errors.