feat(agent): make overload (503/529) backoff configurable

[Ddhjx-code]

Summary

• Exposes three config keys (overload_max_retries, overload_base_delay, overload_max_delay) under agent: so users can tune how aggressively the retry loop backs off on provider 503/529 before falling back or surfacing the error.
• Splits the previously coupled overloaded/timeout fallback threshold so each can evolve independently.
• Defaults match prior hardcoded behavior (2 retries, 2.0s base, 60.0s max) — zero behavioral change when unconfigured.

Closes #55540

Changes

File	What
agent/agent_init.py	Read 3 new config keys with type-safe parsing and floor clamping
agent/conversation_loop.py	Use configured values for overload fallback threshold and backoff delays
cli-config.yaml.example	Document the new keys with usage guidance
tests/agent/test_overload_backoff_config.py	15 tests covering defaults, custom values, clamping, invalid input, threshold logic, and delay behavior

Test plan

• New unit tests pass (pytest tests/agent/test_overload_backoff_config.py — 15 passed)
• Existing output-cap and retry tests unaffected
• Manual: set overload_max_retries: 0 and verify immediate fallback on 503
• Manual: set overload_base_delay: 30.0 and verify longer waits on overload retry

[Ddhjx-code]

Addressed both findings from the review bot on the mirror PR (hashbender#129):

1. Upper-bound clamp — overload_max_retries is now clamped to api_max_retries at parse time (min(overload, api)), so the overload fallback threshold can never exceed the outer retry loop ceiling.

2. Off-by-one fix — Changed >= to > in the fallback guard so that N means exactly N retries before fallback. Values 0 and 1 are now distinguishable.

Added 3 new test cases covering the clamp and boundary behavior. All 18 tests pass.

See commit 9b169e1.

[teknium1]

Thanks for making the fallback/backoff controls explicit.

Problems

• The PR predicates on FailoverReason.overloaded, but current main uses that same reason for the narrow Z.AI Coding HTTP-429 path and intentionally extends its retry ceiling for the adaptive 30/60/90/120-second schedule (agent/conversation_loop.py:3177-3188, agent/retry_utils.py:108-154; commit 1c702aa73). The proposal says 503/529, so the salvage needs to decide and test whether the new settings should affect this 429 path.
• tests/agent/test_overload_backoff_config.py:15-40 copies the production parser, and :115-159 copies the fallback predicate. Those tests do not execute agent/agent_init.py or agent/conversation_loop.py, so they cannot detect a wiring regression. tests/run_agent/test_api_max_retries_config.py:12-28 provides the existing AIAgent-based pattern.
• The new documented settings are not represented in hermes_cli/config.py:990-1021 or the public retry configuration section at website/docs/user-guide/configuration.md:832-840.

Suggested changes

• Rebase the behavior on the current Z.AI-aware retry path, preserve its narrow adaptive policy, and add production-path tests for generic 503/529 plus the Z.AI 429 boundary.
• Register and document the final configuration surface alongside agent.api_max_retries.

Automated hermes-sweeper review.

[teknium1]

This helper duplicates the implementation under test, so every parsing assertion can pass even if agent/agent_init.py stops assigning these attributes. Please construct AIAgent through the config loader (as tests/run_agent/test_api_max_retries_config.py does) and assert the real initialized fields.