fix(gateway): config.yaml wins over .env for agent/display/timezone settings

[teknium1]

Problem

Gateway silently capped at 60 tool-calling iterations per turn despite agent.max_turns: 500 in config.yaml. Telegram/Discord sessions hit _handle_max_iterations at exactly iteration 60 on every long turn. CLI (which reads agent.max_turns directly) was unaffected.

Root cause was a stale HERMES_MAX_ITERATIONS=60 entry left in .env by an old hermes setup run, combined with a not in os.environ guard pattern in the gateway's config→env bridge that let the stale .env value silently shadow the current config.yaml.

Bridge Inconsistency

The config→env bridge at gateway import time (gateway/run.py) was inconsistent across keys:

• max_turns → unconditional overwrite (config wins) ✓
• gateway_timeout, gateway_timeout_warning, gateway_notify_interval, restart_drain_timeout, gateway_auto_continue_freshness → guarded by if X not in os.environ (stale .env wins) ✗
• display.busy_input_mode, display.busy_ack_enabled → guarded (stale .env wins) ✗
• timezone → guarded (stale .env wins) ✗
• security.redact_secrets → unconditional ✓

The except Exception: pass at the bottom of the bridge also swallowed any partial bridge failures silently, meaning a raised exception anywhere in the block would skip subsequent keys and leave .env values in place — including HERMES_MAX_ITERATIONS.

Fix

1. gateway/run.py — drop not in os.environ guards for all agent.*, display.*, timezone, and security.* bridge keys. config.yaml is now authoritative for these, matching the semantics already in place for max_turns, terminal.*, and auxiliary.*. Also surface bridge failures to stderr instead of pass so operators see breakage.

2. gateway/run.py — INFO-log the resolved max_iterations at gateway start so operators can verify the bridge did the right thing instead of chasing a phantom budget ceiling:

INFO gateway.run: Agent budget: max_iterations=500 (agent.max_turns from config.yaml, or HERMES_MAX_ITERATIONS from .env, or default 90)

3. hermes_cli/setup.py — stop writing HERMES_MAX_ITERATIONS to .env in the setup wizard. config.yaml is the single source of truth. Also proactively remove any stale .env entry left behind by pre-fix setups on next run.

Tests

• tests/gateway/test_config_env_bridge_authority.py (new, 166 lines) — regression tests for every bridge key, asserting config.yaml wins even when .env has a conflicting value.
• tests/hermes_cli/test_setup_agent_settings.py (+55 lines) — covers the setup.py HERMES_MAX_ITERATIONS cleanup path.

Verification

After the fix, running gateway logs:

2026-05-02 01:49:39,618 INFO gateway.run: Agent budget: max_iterations=500 (agent.max_turns from config.yaml, or HERMES_MAX_ITERATIONS from .env, or default 90)

Telegram sessions no longer cap at 60 tool calls per turn.