fix: guard float(os.getenv()) against non-numeric env var values

[annguyenNous]

Problem

float(os.getenv('VAR', '1800')) raises ValueError when the env var is set to a non-numeric string like "abc". The or default pattern doesn't help because non-empty strings are truthy — float("abc" or 1800) still raises.

This affects 5 call sites in the core request path (hit on every API call):

File	Env Var	Impact
run_agent.py	HERMES_API_TIMEOUT	Every request timeout resolution
run_agent.py	HERMES_API_CALL_STALE_TIMEOUT	Stale detector base timeout
chat_completion_helpers.py	HERMES_API_TIMEOUT	Streaming request timeout
chat_completion_helpers.py	HERMES_STREAM_READ_TIMEOUT	Stream read timeout
chat_completion_helpers.py	HERMES_STREAM_STALE_TIMEOUT	Stream stale detector

Fix

Add _safe_float_env(var_name, default) helper that wraps float() in try/except (ValueError, TypeError) with a warning log. Applied to all 5 unguarded call sites in the core request path.

Additional unguarded sites exist in gateway platforms (discord, wecom, telegram) and CLI auth — those only affect startup and can be addressed in a follow-up.

Tests

• py_compile passes on both files
• Helper returns default when env var is None
• Helper returns default when env var is non-numeric
• Helper returns parsed value when env var is valid numeric string

[teknium1]

Thanks for identifying the malformed environment-value failure mode. The remaining non-stream stale-timeout path is still live on current main, but most of this PR has since been covered by the shared helper.

Problems

• run_agent.py:1257-1259 still directly calls float(env_timeout), but the other four proposed sites already use canonical env_float after a7dd98c8609c0d944e3c5dd0c5b9ee31dd99eb29 (run_agent.py:1233; agent/chat_completion_helpers.py:2171, 2178, 2998).
• The PR's stale fallback is 300.0, whereas current main's documented and tested default is 90.0 (run_agent.py:1242-1246, 1274; tests/hermes_cli/test_timeouts.py:267-268).
• tests/hermes_cli/test_timeouts.py:228-268 lacks a malformed-value case.

Suggested changes

• Salvage the one live branch with env_float("HERMES_API_CALL_STALE_TIMEOUT", 90.0) and retain its explicit-value flag.
• Add a malformed-env regression case asserting (90.0, False).

Automated hermes-sweeper review.

[teknium1]

When salvaging this branch onto current main, use the canonical env_float helper and preserve the current 90-second default. Current main documents and tests 90.0 for the implicit stale timeout (run_agent.py:1242-1246, tests/hermes_cli/test_timeouts.py:267-268), so this PR's 300.0 fallback would regress that contract.