feat(matrix): state the vocabulary, and let two games have two clocks

[drewstone]

Three faults, all found by running the first two-game study this repo has completed.

A policy was never told what words the game takes

The streaming transport states them in brief.json, exactly so an agent does not have to guess. The persistent and per-decision transports sent frame, history and context and never said.

Measured: a control that cycles 2048's four words scored 1948 there and 0 on ALE Breakout across three replicates, with distinctInputs=1 and no lives lost — Breakout takes NOOP FIRE RIGHT LEFT and none of the four it emitted was legal. Every word was substituted, so it read as a player that did nothing rather than one that was never told the rules.

Both transports now carry commands in the request, and the bundled control reads them instead of assuming. With the words stated it launches the ball and loses five lives in 126 decisions.

assertJoinable refused the study it exists to serve

It rejected "2 protocols" across the whole pool. Two games legitimately have two clocks: a packaged reference pins the clock it was recorded at, so Breakout is derivable only at frameskip 4 while native-2048 runs at 1.

A transfer statistic ranks profiles inside each game and correlates the orders, so a raw score never crosses a game boundary. The check is now scoped per game. Two clocks inside one game is still refused — those rows get averaged and the dominant term disappears into the mean.

The bundled controls were unreachable

They lived under runs/, which is gitignored, so any definition naming them was broken for everyone but its author. They are in examples/policies/ now, alongside the two halves of the study that found all of this.

Result

12 cells, 0 blocked, and the first cross-game statistic this repository has produced:

transfer: tau=0.000 over 2 games (1 pairs), 2 of 2 arms distinct
— one game pair: this tau is a single observation, not evidence that an order transfers

The machinery works end to end and the study is uninformative, which are different things. Both controls score 0 on Breakout, so there is no order to correlate; the runner says so rather than presenting the zero as a finding. Separating profiles there needs agents, not cheap controls — neither control can hit a ball.

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