feat: sync moteur 61021f0 vers OSS (2026-08-18)

[Axellum]

Summary

• Sync du miroir public depuis moteur_agents @ 61021f0 (197 commits depuis le 13/08).
• Côté vocal/domo : cascade Zero-LLM (commande → état → météo → date → chat), fast-path Cerebras gpt-oss-120b, sonde d'hôte local, température par pièce, volets unknown sans capteur, marqueurs accentués.
• Exemples à jour : .env.example (privacy, concurrence, HA_WEATHER_ENTITY, fast-path), config.example.json (tiers + budget + routing), ha_commands.json.example (entités living_room / bedroom).
• Anonymisation : IP 192.168.0.x → 192.168.1.10, IDs matériels (Daikin/Sonoff/Tuya) → slugs génériques, commune météo → weather.home, pas de token / .env / config.json.

Test plan

• python scripts/oss_secrets_scan.py exit 0
• Tests vocaux/HA locaux verts (état, météo, volet, cascade, fast-path sonde, AJEAN local)
• CI 🔒 OSS secrets scan + 🧪 Pytest sur cette PR

---

Note

High Risk
Wide behavioral changes to execute/stream concurrency, HITL bypass for autonomous tasks, and acceptance-contract verdicts directly affect production safety and session lifecycle; misconfiguration of concurrency or contract rules could block users or merge bad work.

Overview
Large mirror sync that reshapes how /api/execute and background runs share the engine: a per-session execution registry with MOTEUR_EXECUTION_MAX_CONCURRENCY, duplicate-session 409s without burning LLM routing, and heavy pipelines taking slots only at run_full_pipeline while chat/HA fast paths and mode=chat stay outside the registry (#T324/#T344). Stop/abort/stream paths now clear or release registry entries instead of a single global lock.

Acceptance contracts and review get stricter after production incidents: planner can declare traced criterion substitutions; initial file state is snapshotted so “already true” criteria cannot absolve a failed DAG; the engine keeps cause_echec, reads the first persisted contract for HITL resume, and task_ sessions bypass blocking HITL like other autonomous night jobs.

DreamCoder uses real pipeline/envelope timeouts, best-effort git push of task/* branches (with cycle reporting), and fixes backlog tasks stuck on failed with retries=0. DAG task watchdogs become configurable (base + per-tier). Executor refuses tool calls with invalid JSON instead of running tools with {}. OpenAI proxy omits or relays real token usage only (#T357). HA fuzzy deprioritizes unavailable/unknown entities.

Examples and ops docs shift to generic deploy values (vromvrom-engine, placeholder LAN IP, privacy/concurrency env vars, richer config.example.json); health exposes short commit on /healthz and full version on authenticated observability; models-health adds catalogue “grey zone” audit; Dashscope drops from budget rotation.

^{Reviewed by Cursor Bugbot for commit 85536f8. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: serverGenReqId_f4f7c639-cdb0-47cf-8dcf-534440d70868)