fix: emergency compression before max_iterations exhaustion

[gzsiang]

Problem

When max_iterations (default 90) was exhausted before the compression
threshold was reached, the agent was killed without compressing.

This happened when:

• Provider doesn't return usage data (last_prompt_tokens == 0 → should_compress(0) never fires)
• Tool outputs are small per iteration (context accumulates slowly, never reaches threshold in 90 iterations)

Result: Agent hits max_iterations, calls _handle_max_iterations, and the session ends mid-work.

Fix

Add emergency compression in the main loop that triggers when:

• iteration budget is nearly exhausted (remaining <= 1)
• compression is enabled
• estimated message tokens exceed the compression threshold
• fewer than 3 emergency compressions this turn

After compression, the budget resets and the agent continues.

Changes

• IterationBudget.reset() method
• Emergency compression check before api_call_count += 1 in the main loop
• _emergency_compression_count counter (reset at turn start)

[liuhao1024]

The emergency compression mechanism is a solid idea for handling the "provider doesn't return usage data" edge case. Two concerns:

1. No tests for a complex behavioral change

This PR adds a new IterationBudget.reset() method and a new code path that can extend the agent's effective iteration budget up to 4x max_iterations (original + 3 resets). The diff only touches run_agent.py — no test file. Given the interaction between _emergency_compression_count, iteration_budget.reset(), api_call_count reset, and the _budget_grace_call gate, this deserves at least:

• A unit test for IterationBudget.reset() (reset mid-consumption, verify remaining and used)
• An integration test showing that when remaining <= 1 and context is above threshold, emergency compression fires and the budget resets
• A test verifying the cap at 3 emergency compressions (4th exhaustion should hit _handle_max_iterations)

2. #TODO-issue placeholder in the comment block (line ~10718)

The comment has (#TODO-issue) — should reference the actual issue number before merging.

[teknium1]

Thanks for the contribution. Current main already implements the requested continuation-safe compression behavior, so this PR is now redundant.

• agent/conversation_loop.py:1007-1058 evaluates current request pressure before each model call, compresses at the canonical threshold, then refunds the iteration and continues.
• tests/run_agent/test_run_agent_codex_responses.py:1505-1556 covers a tool-heavy mid-turn compression followed by successful completion.
• Commit a6e0ad955ec59d99b91b804ce0a3d2f6ca4cf5c1 explicitly records integration of fix: emergency compression before max_iterations exhaustion #18607; the current path was subsequently refactored and persistence-hardened in 1f430e1aa23c8ace76c3506df5005be0edd5b744 and 475dd972636904a330f0088f2f5c1e9ee2cc9eb4.

This is an automated hermes-sweeper review.