Add the complete three-watch Bellwether human-play game

[concordia-claw]

Consolidated into #372. Closed only as a superseded review thread, following Joel Z. Leibo’s category guidance. The work is preserved in the continuing category PR; this does not indicate an upstream merge or rejection of the feature. Continuing issue: #371; continuing PR: #372. Reusable model-adapter/prefab additions continue in PR #356; the Bellwether example continues in PR #372.

---

Closes #317.

Dependencies / review scope

Dependent draft; do not merge before #316. That prerequisite in turn includes #306 (human input and prefab injection from #302/#304) and #312 (lifecycle). This branch continues #316's exact reviewed head 6a006731; no prerequisite has been merged into upstream.

Feature-only comparison — 16 files, three focused commits:

1. cd7c597: optional timeout/output bounds in the existing Ollama adapter; honors requested text limits, with mocked request tests. Separate from feat: expand markdown sanitization to Groq, Ollama, and HuggingFace wrappers #260's markdown sanitization proposal.
2. d93ba2a: extract minimal's existing optional-context assembly and reuse it in basic. Without this, basic silently ignores the resident account/affiliation components. Default basic/minimal policies and insertion semantics stay unchanged.
3. f59e142: full Bellwether game, player experience, configuration, documentation and tests.

What this adds

• Dusk → High Tide → Before Dawn → epilogue, four consequential human choices per watch. Free map/cast inspection; public/private discussion and open-ended proposals; explicit consent, promises, allocations, transfers, repair orders and waiting.
• Example-owned StormNight resolution and standard Inventory.apply accounting. Generator 6 + owned reserve 2 versus 9 baseline facility demands. Part + accepted and performed Ivo labor before the final watch reduces beacon demand. Dialogue never creates resources, consent or work.
• Standard Simulation/Sequential, SwitchAct routing, observation queue, memory, minimal/basic resident prefabs, and JSON/HTML logs. Four live residents can refuse or revoke; their policies are not human or forced-consensus fixtures.
• Server-filtered disputed-account recipients and private dialogue; institutional rules, knowledge, enforcement, commitments and actual consequences recorded separately.
• Extend Add shared attached operations and Bellwether human-play fixture #316's shared operation service/player page, not a second engine/server. New game.begin, game.preview, and run.resume operations share the existing GUI/CLI path. Live workers—including paused workers waiting for input—still cannot be edited. Reload does not restart.
• --mode slice remains the original one-turn fixture. --mode fixture is a clearly labelled cooperative full-night fixture. --mode live --model llama3.2:3b uses bounded local Ollama actors. See concordia/examples/bellwether/GAME.md for installation, both strategies, exact command vocabulary, --resident-prefab basic, and --dispute-file.

Verification actually performed

• 62 affected checks pass (Bellwether rules/service/privacy/Chromium, existing prefab/default-injection regressions, and mocked Ollama bounds). Both advertised resident configurations retain their standard contexts, private account, affiliations and normal ConcatAct policy.
• Type checking: 0 errors, existing scoped suppressions retained. No new pylint findings; 10 pre-existing long-line findings remain in the two prefab files. Contribution formatting/imports and diff whitespace checked. Wheel built with game modules, player page and guide.
• Actual Chromium + attached CLI against the standard server/Sequential after explicit launch authorization:
  • Cooperative fixture: 12 human choices / 24 engine steps, 9/9 facility-watches using 8 fuel, accepted/performed repair.
  • Prioritization fixture: 12 choices / 20 steps, 6/9 using 6 fuel, reserve untouched, no repair; materially different grounded consequences.
  • Local live llama3.2:3b, standard minimal actors: 12 choices / 24 steps, four resident dawn responses. Nell declined; no consent/transfer/repair was invented. 6/9 services, including recorded failed service promises.
• All three browser journeys had zero page errors, mobile overflow checks, reload/reconnect recovery, free inspection, clarification without effects and attached CLI state checks. Committed browser regression also checks literal text, private routing and GUI/CLI retry equivalence on synthetic input (not another simulation).
• Live standard profiler: 12 successful local calls, 2.936–6.013s, mean 4.344s; longest group response 19.306s and final dawn responses 17.705s. Per-click browser timings do not include later resident work and are not used as full-action latency. Cost unmeasured/null; no paid model calls.

Explicit limits

This is a playable bounded game, not completion of the broader authoring/checkpoint/experiment showcase. No checkpoint restoration/isolation claims, general institution editor, CRUD, undo/redo, cancellation of live model work, multiplayer, or automatic restart. Original broad six-workflow baseline and A–H showcase completion remain unchanged.

Natural-language discussion/proposals stay verbatim. Material operations use a conservative documented vocabulary; ambiguous/unsupported physical attempts ask for clarification before effects instead of guessing. One local live trajectory is not narrative-quality, human-usability, social-scientific, or deterministic-reproduction evidence. A 15–25 minute human playthrough and live basic-policy trajectory were not measured (basic composition is regression-checked). Private verification processes were stopped; no existing game/routes changed.