fix(feishu): adapter declares enforces_own_access_policy so gateway honors group_rules

[hashbender]

Summary

Feishu gates DM/group access at intake via FEISHU_GROUP_POLICY + per-group group_rules.<chat_id>.allowlist (parsed from config.extra). The gateway's env-based FEISHU_ALLOWED_USERS check runs AFTER this.

Problem: when no FEISHU_ALLOWED_USERS env allowlist is configured, the gateway had no way to honor the config-driven group_rules allowlist the adapter already enforced — it default-denied every Feishu group message regardless of group_rules. This is because FeishuAdapter did not declare enforces_own_access_policy = True (unlike WeCom / Weixin / Yuanbao / QQBot / WhatsApp), so the gateway's _is_user_authorized did not trust the adapter's intake-layer decision.

Fix: add enforces_own_access_policy property returning True on FeishuAdapter. The gateway now trusts the adapter's decision when chat_type is group, so group_rules becomes sufficient at the gateway layer without requiring a parallel FEISHU_ALLOWED_USERS env allowlist.

Root cause

Two-layer gating gap:

1. Adapter layer (_allow_group_message in plugins/platforms/feishu/adapter.py): correctly uses group_rules[chat_id].allowlist — works.
2. Gateway layer (_is_user_authorized in gateway/authz_mixin.py): checks _adapter_enforces_own_access_policy(platform) via getattr(adapter, "enforces_own_access_policy", False). Feishu returned False → gateway skipped the trust-adapter branch → default-deny for all Feishu group messages.

Changes

• plugins/platforms/feishu/adapter.py: add enforces_own_access_policy property (returns True) to FeishuAdapter. Mirrors WeCom / Weixin / Yuanbao / QQBot / WhatsApp.
• tests/gateway/test_config_driven_access_policy.py: extend Feishu coverage — _OWN_POLICY_PLATFORMS, _clear_auth_env (FEISHU_ALLOWED_USERS / FEISHU_ALLOW_ALL_USERS), test_own_policy_adapters_declare_the_flag.

Verification

./venv/bin/pytest tests/gateway/test_config_driven_access_policy.py tests/gateway/test_multiplex_profile_authz.py -q
# 85 passed

Checklist

• Patch only touches the two relevant files
• No new core tools / model tool schema changes
• Prompt caching unaffected (no system prompt / tool schema change)
• Mirrors existing pattern (WeCom et al.) — no new API surface

---

Mirror-of: NousResearch#56838
NousResearch#56838