fix(gateway): keep max iterations config authoritative

[Soju06]

Summary

• Keep gateway max_iterations resolution authoritative to config.yaml agent.max_turns when it is configured
• Reload .env before resolving the gateway agent budget so stale HERMES_MAX_ITERATIONS values cannot win during gateway turns
• Cover the stale .env regression with a focused gateway test

Why

A stale HERMES_MAX_ITERATIONS=90 entry in .env can cause gateway sessions to stop with Iteration budget exhausted (90/90) even when config.yaml sets agent.max_turns higher. Removing the .env line is enough as an immediate local mitigation, but the gateway should preserve the existing config-authority contract so stale regenerated env values do not reintroduce the bug.

Test Plan

• python -m pytest tests/gateway/test_runtime_env_reload_config_authority.py tests/gateway/test_run_cleanup_progress.py -q
• python -m compileall -q gateway/run.py
• git diff --check upstream/main..HEAD

[liuhao1024]

Verification Review

Reviewed by: automated PR review cron

Summary: Clean refactor that makes config.yaml the authoritative source for across all three gateway agent-creation paths. The extracted and helpers eliminate the duplicated config-reading logic and ensure stale values cannot override explicit config.

Observations:

• The precedence chain (config.yaml → .env → default 90) is correctly enforced in all three call sites
• now only bridges to env when config explicitly sets the value — previously it would set the env var even from a key that was absent, leaving stale .env values in place
• Test coverage verifies the config-overrides-env-after-reload scenario

LGTM.

[kshitijk4poor]

Closing as already resolved on current main.

The config-authority contract this PR restores is already in place in gateway/run.py:

• _reload_runtime_env_preserving_config_authority() reloads .env for fresh credentials but explicitly preserves config.yaml's agent.max_turns so a stale HERMES_MAX_ITERATIONS in .env can't override it on later turns.
• _bridge_max_turns_from_config() bridges config.yaml agent.max_turns into the env var, and the startup bridge guards the "stale 60 vs configured 500" regression.
• _current_max_iterations() re-reads the budget per turn after that refresh.

So a stale HERMES_MAX_ITERATIONS=90 no longer wins over a higher configured agent.max_turns. Thanks for the fix and the clear write-up of the regression — the contract you wanted is honored on main.