fix(auxiliary): floor config-derived compression timeout at 300s (#54915)

[teknium1]

Summary

Codex/reasoning-model compression summaries no longer die at the 120s auxiliary timeout and silently degrade to a context marker: config-derived compression timeouts now get a bounded 300s floor. Fixes #54915; second half of the "gpt-5.5 compacts continuously near 272K" community report (with #59804).

Root cause: with auxiliary.compression.provider: auto, compression rides the main runtime. Codex gpt-5.5 summarizing a large context legitimately exceeds the default 120s stream deadline; the timeout aborts the summary and the compressor inserts a deterministic marker that barely shrinks the session, so compaction immediately re-triggers.

Salvaged from PR #54943 by @Tranquil-Flow (cherry-picked, authorship preserved). Honors all four issue constraints: bounded (not infinite), compression-task-only, explicit per-call timeout= never floored, higher config values kept. Covers sync + async paths; verified the floor propagates into the Codex Responses stream total_timeout.

Changes

• agent/auxiliary_client.py: _COMPRESSION_TIMEOUT_FLOOR_SECONDS = 300.0 + _effective_aux_timeout() helper; wired into both call_llm and async_call_llm
• tests/agent/test_auxiliary_compression_timeout_floor.py: new suite exercising real call paths with mocked client (floor applied / explicit override honored / other tasks untouched / higher config kept, sync + async)

Validation

	Before	After
compression, config timeout 120s	120s reaches client	300s
explicit timeout=60 on compression	60s	60s (unchanged)
title_generation, config 30s	30s	30s (unchanged)
tests: timeout_floor + auxiliary_main_first	—	29/29 pass

E2E: real call_llm(task="compression") with mocked openai-codex client asserts timeout=300.0 reaches completions.create.

Infographic