v0.1.0 — family-diverse AI playtesting

First public GitHub release. Dogfood swarm #2, 2026-09-14. 84 -> 182 tests,
then the Law 8 execute set. First live run against claude-rpg (proof-01,
8 turns, mistral). Stage B: 24 HIGH. Stage C: remaining MED/LOW
(humanization — errors that name the cause, CLI help, report honesty).
Feature execute: consume Observation.actions, serial RPC reset(),
viewport-only PTY grid, report legends, schemaVersion on leaving
artefacts, single-run REPORT.json.

Not on npm (package.json stays "private": true).

Added

• Typed actions from Observation.actions. Player replies map onto
  choose / key / line. A closed-set miss is a harness event
  (illegal-action); the runner does not step. describeActions() is appended
  to the player prompt.
• Serial RPC reuses one driver. --serial on driver.kind === "rpc"
  starts once, calls reset() between seats, and stops once. A missing or
  throwing reset fails that next seat with E_RESET (no silent second
  client). Parallel RPC stays one process per seat.
• PTY viewport grid. grid.lines is viewportY .. viewportY+rows-1, not
  the scrollback buffer. Prompt matching uses baseY+cursorY. kind:key
  writes without a trailing CR.
• Single-run REPORT.json sidecar (kind: "single-run-report") next to
  REPORT.md. Markdown legends teach !, repeat, loop, and H(a) in
  bits. schemaVersion + reportFormat + VERSION are stamped on the
  markdown generator line, both JSON reports, and seat meta.json.
• State-gated invariants. HP never negative and inventory non-decreasing,
  only when a turn carries a structured state object — never inferred from
  prose.
• Landing page + Starlight handbook (site/, accent cyan). Pages workflow
  path-filtered to site/**. Live at the org GitHub Pages URL.
• README translations (ja, zh, es, fr, hi, it, pt-BR) via TranslateGemma 27B.
• README lockup from mcp-tool-shop-org/brand (logos/ai-playtest/readme.png).
• Deterministic transcript verifiers (src/verifiers.ts). Six checks, all
  transcript-only: absorbing-SCC (Tarjan over the observed screen digraph,
  kind=review, never a trap proof), ignored-input attribution, parser
  unparsed/refused/accepted with empty defaults, terminal victory/death,
  no-progress windows, entity-appearance leads for the jury. Specified in
  docs/research-2.md §D; false-positive modes printed beside every hit.
• --runs N with a Beta-Binomial posterior Beta(s+1, n−s+1), stability
  labels STABLE_PASS / STABLE_FAIL / UNSTABLE, worst-of-n beside the mean,
  and the copy that n=3 is descriptive (2/2^n = 0.25). First n that can
  clear α=0.05 is 6. Do not bootstrap.
• Jury n_eff on the report: k / (1+(k−1)·φ̄), warn when n_eff/k < 0.5.
• RPC hello {protocol:1} is a hard handshake. A bridge that only
  speaks observe / act / quit now fails closed on connect. Serial
  runAll now calls reset() between RPC seats.
• stdio game driver (stdio-game.ts) — spawns a game as a child process and
  decides when the game is waiting for input: a promptPatterns regex matched
  the stripped tail and output went quiet for promptQuietMs, or nothing matched
  and idleQuietMs of silence passed. The runner never parses the game.
• OpenRouter seats (openrouter.ts) — one HTTP seam, injectable fetch, six
  retries with exponential backoff (126 s total) for bursty upstream rate limits.
• Model players (player.ts) — a seat sees the screen since its last input
  plus its last playerMemoryTurns exchanges, is briefed by persona (goals and
  register, never the mechanics under test), and answers with one line.
• Scripted setup (config.ts, run.ts) — prompts matching a setup[].match
  regex are answered from the config without spending a turn, so every family
  starts from the same character.
• Critic (critic.ts) — reviews a transcript against the game's criteria[]
  at temperature 0 and returns one JSON object: per-criterion met/evidence/turn,
  an alive verdict, highlights, dead spots, confusions.
• Report (report.ts) — aggregates every seat: criteria by family, verdict
  counts, and every dead spot and confusion named with the seat that found it.
• CLI (cli.ts) — run and report verbs; exit codes 0 ok, 1 usage,
  2 config, 3 provider, 4 run error.
• Config validation refusing two seats from one model family.

Changed

• The runner consumes actions / state. Illegal-action rejection and
  hp/room hashing are wired; docs/engine-bridge.md no longer says they are
  protocol-only.
• panelSize default 1, not 3. Kohli 2026 (Kish n_eff 2.18, panel 72.0% vs
  best single 71.8%, cross-family φ 0.389 vs same-family 0.437) undercuts
  family diversity buys independence. Author-off-jury stays (Panickssery /
  Stechly). Extra jurors flag disagreement. Decision written in
  docs/research-3.md.
• Coverage "thin" on a short-but-varied session is labelled a sample-size
  limit, not "explored thinly / never reached the content" (proof-01).
• Criteria-table agreement uses juror splits on a 1-seat run (proof-01 hid
  1/3 splits behind no! and an em-dash).
• Jury dead spots and confusions appear in the report, not only the author's.
• Three drivers behind one observation seam. stdio (the original),
  pty (a rendered terminal grid, for full-screen TUIs), and rpc (structured
  state over TCP, for Godot / Unreal / anything instrumented). Selected with
  driver in the config; defaults to stdio, so existing configs are
  unaffected. docs/engine-bridge.md carries a paste-and-go Godot 4 autoload.
• A cross-family jury. Each transcript is judged by author-off seats
  (panelSize, default 1). Majority verdict, split criteria marked, dispersion
  reported. (Shipped at default 3 in swarm #1; default moved to 1 in swarm #2 —
  see Changed.)
• Coverage. Novelty curve and half-life, repeat / loop / self-loop rates,
  action entropy and a thin/moderate/broad read, computed from turn records
  alone.
• LICENSE (MIT), SECURITY.md with the threat model, CI, prepublishOnly,
  tsconfig.test.json, and this changelog.

Fixed

• Last player input's resulting screen was labelled quit. The quit loop
  consumed it, so the critic never saw the last action's result (proof-01:
  /director opened director mode; the recap the jury cited was the save
  screen). Consequence is now recorded first; runner quit inputs stay quit.
• Verdicts could be inverted. parseCritique coerced with Boolean(), and
  Boolean("false") === true, so a critic answering "alive": "false" as a
  string was recorded as alive with every criterion met.
• The OSC branch of the ANSI regex over-deleted. It terminated on the next
  BEL anywhere in the buffer, so an ST-terminated OSC plus a later BEL removed
  everything between them — 46 characters of narration in the measured case.
  That regex also stored literal control bytes, so reviewers saw a different
  pattern than the one that ran.
• OPENROUTER_API_KEY reached the game process and was observed rendered
  into screen text, from there into the player's context and the written report.
  The child now gets an allowlist; game.inheritEnv opts back in.
• The critic's evidence included the runner's own quit inputs as player
  decisions and excluded every terminal screen, so crashes and stalls were
  invisible to the verdict.
• The transcript sat last and unframed in the critic's prompt — the most
  instruction-weighted position — so a game could steer its own grade.
• stderr shared a quiet-timer with stdout (3063 ms timeout every turn where
  220 ms prompt was correct); stdout chunks were decoded independently,
  corrupting split multi-byte characters; an unhandled EPIPE on stdin took down
  every parallel seat; a failed spawn reported nothing.
• finish_reason: 'length' was returned as success, and res.text() sat
  outside the retry loop's try/catch.
• --label ../../etc escaped runsDir; --turns abc played zero turns and
  exited 0; Promise.all discarded every sibling when one seat threw; a run
  where every critique failed exited 0; the report read "Alive verdicts: 1 of 1"
  for a two-seat run with one dead seat.
• Critic token budget raised 1,800 → 6,000. Nineteen criteria with evidence
  strings no longer fit in 1,800 tokens; one seat's critique was cut mid-array
  and failed to parse twice. Sized for roughly 25 criteria. Retries now carry the
  previous parse error so the second attempt is informed.