Problem

Closes #36718.

After compress() runs, the scheduler sets:

• last_prompt_tokens = -1 (sentinel)
• awaiting_real_usage_after_compression = True

But last_real_prompt_tokens still holds the old pre-compression value (above the threshold). On the very next preflight check, should_defer_preflight_to_real_usage() hit this branch:

if self.last_real_prompt_tokens >= self.threshold_tokens:
    return False   # incorrectly skips deferral

…returned False, allowing should_compress(preflight_tokens) to fire — triggering a second compression before the provider ever reported real token usage for the now-shorter conversation.

Root cause

awaiting_real_usage_after_compression exists precisely to guard this window, but should_defer_preflight_to_real_usage() never consulted it. The stale last_real_prompt_tokens value short-circuited deferral.

Fix

Add an early-return in should_defer_preflight_to_real_usage() (agent/context_compressor.py):

if self.awaiting_real_usage_after_compression:
    return True   # defer until update_from_response() clears the flag

update_from_response() already clears awaiting_real_usage_after_compression once real prompt_tokens arrive from the provider, so the guard is active for exactly one turn.

Test

Added two cases to TestPreflightDeferral in tests/agent/test_context_compressor.py:

1. test_defers_immediately_after_compression_before_real_usage_arrives — verifies should_defer_preflight_to_real_usage returns True when the flag is set and rough tokens exceed threshold (the previously broken case).
2. test_no_longer_defers_after_real_usage_clears_flag — verifies normal baseline/growth deferral resumes once the flag is cleared.

Test plan

• Existing TestPreflightDeferral tests still pass
• TestUpdateFromResponse tests still pass (flag-clearing path unchanged)
• Manual: start a long conversation that triggers compression; verify only one compression fires per threshold crossing, not two back-to-back