chore(release): 0.10.0

[drewstone]

Seven commits since v0.9.0. The changelog entry states each defect with the measurement that found it, in the format 0.9.0 set.

What 0.10.0 is

An arena. One definition becomes cells; a cell is one profile, one game, one objective, one protocol, one sensor, repeated reps times. Every cell is played, every replay attested, and the summary is mean pairwise Kendall tau-b between per-game rankings.

Three transports, because the choice outweighs the model. Same program on native-2048: score 4 under one process per decision, score 1948 under one process per episode. Per-decision spawning silently forbids state while still answering every decision and attesting clean.

Authoring. Across nine cells the rank correlation between moves delivered and score was 0.94 — a live transport very nearly measures typing speed. author= gives the agent a practice game at a different seed and a build budget, then runs the program it leaves cold against a fresh scored instance. Measured, two replicates each:

profile	score	decisions played	build tokens
opus	46,124 / 46,216	2000 / 2000	1.19M / 0.96M
sonnet	8,944 / 3,888	609 / 317	5.14M / 2.76M
haiku	4 / 60	2000 / 2000	2.23M / 1.96M

Opus's runs differ by 92 points on 46,000. Sonnet spent 4.3x the tokens on a program that filled the board and lost in under a third of the horizon. The same model typing live scored 240.

Seven numbers that were not what they said are listed in the changelog, including a row that scored a hardcoded channel so any game not publishing score could never be ranked, a rank correlation that would have reported a perfect transfer as a perfect inversion, and a practice game that wrote 449,821 files and 1.7 GB because pace=0 is correct for a scored episode and wrong for practice.

Composition, not reimplementation.improve already owns a shot loop over ten profile surfaces and compareOptimizationMethods already owns disjoint splits with Bonferroni-adjusted intervals. Playproof supplies a scenario and a score. examples/agent-eval-fit.mts exists to be compiled so the join cannot rot silently.

Gates

Local: 20 of 20 test files, boundary, attestation-blindness, 0 typecheck errors. The real-emulator gates run in CI. pnpm test:pyboy with the Tetris ROM is the release manager's step before tagging, per docs/releasing.md.

One caveat for the tagger: the RetroArch same-process replay gate failed four times on 2026-08-26 and this branch carries a fix for a silent frame-advance barrier that is a diagnostic, not a proven cause — COMMAND_TIMEOUT is 5s with resends, so a timeout means wedged rather than busy. It has passed once since. Watch that job.