fix: dead regex in quotaResetDelay, write_count race, env info lock

[aaronlab]

Summary

• run_agent.py: Fix double-escaped \d in quotaResetDelay regex (line 2313) — the pattern \\d in a raw string matches a literal backslash+d, never matching actual digits. This makes rate-limit reset time extraction dead code. The companion sec_match regex (line 2320) uses \d correctly, confirming this is a bug, not a style choice.
• hermes_state.py: Move _write_count += 1 inside self._lock to prevent concurrent _execute_write callers from racing on the counter, which could skip or double WAL checkpoints.
• tools/terminal_tool.py: Acquire _env_lock in get_active_environments_info() before reading _active_environments. Without the lock, a concurrent cleanup_vm() call can modify the dict mid-iteration, raising RuntimeError: dictionary changed size during iteration.

Test plan

• Verify _extract_api_error_context correctly parses quotaResetDelay: 30000ms from Gemini API errors
• Confirm WAL checkpoints fire at consistent intervals under concurrent writes
• Confirm get_active_environments_info() no longer raises under concurrent environment cleanup

🤖 Generated with Claude Code

[alt-glitch]

Related to #6655 — overlapping fix for _env_lock in terminal tool shared dict accesses.

[alt-glitch]

Related to #6655 — overlapping fix for _env_lock in terminal tool shared dict accesses.

[teknium1]

Thanks for identifying the write-counter and environment-registry concerns. The current tree has moved since this branch, so this needs a focused salvage rather than a direct apply.

Problems

• The tools/terminal_tool.py hunk targets get_active_environments_info(), removed by 8d023e43; it cannot protect the live registry paths. Current unlocked accesses remain at tools/terminal_tool.py:1550-1552, :1669, and :1765-1771. The related fix(terminal): acquire _env_lock for all shared dict accesses #6655 covers this broader class.
• On current main, _write_count drives both checkpoint and FTS-optimize cadence at hermes_state.py:1172-1176. Porting only needs_ckpt would leave the optimize decision outside the counter lock.
• No tests were added for the quota-delay parser or concurrent counter maintenance.

Suggested changes

• Compute both maintenance decisions under the same self._lock acquisition, then run maintenance after releasing it.
• Drop the already-implemented regex hunk (agent/agent_runtime_helpers.py:3062) and retarget the terminal work to current lifecycle functions.
• Add focused regression tests for the retained behavior.

Automated hermes-sweeper review.