fix(auxiliary): floor compression timeout so reasoning models don't fall back to marker (#54915)

[Tranquil-Flow]

Problem

When Hermes uses a reasoning model (e.g. Codex / GPT-5.5) as the auxiliary
compression provider (auxiliary.compression.provider: auto), context
compression can time out at the default 120 s and silently fall back to the
deterministic context marker — losing the LLM summary:

Compression summary failed: Codex auxiliary Responses stream exceeded 120.0s total timeout.
Inserted a fallback context marker.

The root cause is that call_llm(task="compression", ...) resolves its
timeout from auxiliary.compression.timeout (default 120 s) with no floor.
A reasoning model summarising a large context can legitimately exceed 120 s.

Fix

Adds a bounded 300 s floor to config-derived compression timeouts only:

• New constant _COMPRESSION_TIMEOUT_FLOOR_SECONDS = 300.0 (finite, bounded).
• New helper _effective_aux_timeout(task, timeout) — single source of truth
  used by both call_llm (sync) and async_call_llm (async). It resolves the
  timeout as before, then applies max(effective, floor) only when
  timeout is None and task == "compression".

This honours all constraints from the issue:

• Only compression is floored — other auxiliary tasks keep their own timeouts.
• An explicit per-call timeout= override is not floored.
• The floor is a minimum (max), so a higher config value is kept unchanged.
• The timeout stays bounded (never infinite).
• Both sync and async paths are covered.

Verification

• RED: 2 tests fail on upstream/main f1345290e — the 120 s config timeout
  reaches the client unfloored (assert 120.0 >= 300.0).
• GREEN: all 7 new tests pass with the fix.
• Regression: 223 passed across the auxiliary + compressor test suites.

Tests exercise the real call_llm / async_call_llm production paths (mocked
LLM client) and assert the timeout that reaches client.chat.completions.create.

Closes #54915.

---

Auto-published by Moonsong via Path B automated pipeline.

[teknium1]

Merged via PR #59807 — your commit was cherry-picked onto current main with authorship preserved (370a489). Verified the floor propagates into the Codex Responses stream total_timeout (the exact failing path from the issue). Thanks!