fix(ooda): machine-checkable done-gate unsticks the coin-gym goal (UNCLEAR-CRITERIA)

[rysweet]

Triage & course-correction

The standing goal `build-a-local-coin-benchmark-harness-and-a-self-09e65e35`
("build a local coin benchmark harness and a self-improvement loop") was durably
`Blocked` with why=UNCLEAR-CRITERIA — "no tracked PR/issue the done-gate can
verify."

Root cause: not technical. The LOCAL COIN Gym harness and its Phase-5
self-improvement loop are already built and green (src/bin/coin_gym.rs,
src/coin_gym/, gadugi scenarios, docs). The goal re-stalled every cycle only
because its finish condition had no machine-checkable definition — so each
OODA Observe tick re-saw it as unfinished and emitted NO ACTION.

Course-correction chosen (option a — make the done-gate machine-checkable),
mirroring the #4210 / kgpacks-parity fix:

• Specs/coin-gym-benchmark-harness.md — binds the goal's finish condition
  to a single verifiable command and enumerates seven measurable done-criteria
  (target-loader, baseline/team runners, scorer, leaderboard-comparator,
  self-improvement-loop, contract-wiring), each asserted by the harness's own
  coin-gym verify self-check.
• scripts/check-coin-gym-done-gate.sh — one command the done-gate can run
  each cycle; builds coin-gym and runs coin-gym verify, exiting 0 only when
  all seven LOCAL criteria pass. --full also runs the two outside-in gadugi
  scenarios.
• docs/operations/coin-gym-goal-signal-2026-07-18.md — the plain-language
  operator/Signal notification (linked from mkdocs nav).

Verification

$ scripts/check-coin-gym-done-gate.sh
  [PASS] target-loader            5 pinned + 2 held-out-fresh target(s); both families present
  [PASS] baseline-runner          5 outcome(s) for 5 pinned target(s)
  [PASS] team-runner              5 outcome(s) for 5 pinned target(s)
  [PASS] scorer                   reach 60.0% / precision 60.0%; 2 family split; histogram covers 5/5
  [PASS] leaderboard-comparator   compared vs published 'GPT-5.4' (reach Δ +37.1 pts)
  [PASS] self-improvement-loop    held-out reach 0.0% → 100.0% (kept 2, rolled back 0)
  [PASS] contract-wiring          evaluate (8 args) + verify (4 args) argv present; LOCAL-ONLY=true
result: 7/7 criteria passed
✅ DONE — exit 0

Live VM grading (coin evaluate/verify, issue #2823) is Phase 3 and stays out
of this LOCAL done-gate. No harness behaviour changed — this only gives the
done-gate a command it can run so the goal is certified complete instead of
re-stalling.

Operator Signal message (plain English, as sent)

Quick update on the "local coin benchmark harness" goal — good news, it's
actually already finished. The local benchmark and its self-improving loop are
built and every one of their built-in checks passes (7 out of 7). It kept
showing up as "stuck" only because there was no automatic test tied to it that
could confirm it was done. I've added that test now — a single command that
runs the benchmark's own self-check and passes cleanly — so the system can
confirm this goal is complete on its own and will stop reporting it as stuck.
Nothing needed from you.

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com