fix(compression): break infinite loop when aux model context < session size (#53008)

[Christopher-Schulze]

Summary

Fixes #53008. When Hermes uses an auxiliary compression model with a smaller context window than the main model, compression could repeatedly re-trigger without producing enough headroom.

Root cause

• Feasibility correction set the compression threshold to 100% of the auxiliary model context, leaving no room for the summary and protected conversation tail.
• Once the actual middle window exceeded the auxiliary context, summary generation still selected the undersized auxiliary runtime.

Fix

1. Auto-corrected thresholds now use 80% of auxiliary context, clamped to MINIMUM_CONTEXT_LENGTH.
2. compress() compares the actual middle window with auxiliary context. Oversized windows use the main runtime for that pass, then restore the configured auxiliary model in finally.
3. The fallback passes the main provider, model, endpoint, API key, and API mode explicitly, so call_llm(task="compression") cannot re-select the auxiliary task configuration.
4. The overflow warning is emitted once per session and reset on both session reset and natural session end.

Current main already contains provider-token anti-thrash verification. This rebase deliberately removes the PR's older threshold-escape implementation instead of duplicating that upstream behavior.

Verification

• scripts/check.sh --project hermes-agent --worktree worktrees/hermes-agent/53008: PASS
• Locked dependency sync, uv lock --check, and full ruff check .: PASS
• Changed tests: 32/32 PASS
• Focused compression integration set: 47/47 PASS
• Exact-base proof on 7b5ba2054721dde998ed47fd4a0f031955278e99: 6 regression failures on base, all pass on this branch
• git diff --check origin/main: PASS

Full-project ty remains advisory and hits its existing upstream panic in tools/checkpoint_manager.py; the blocking project gates pass.

[teknium1]

Thanks for the focused compression investigation and regression coverage.

This is an automated hermes-sweeper review. Current main already provides the requested infinite-loop guarantee:

• agent/context_compressor.py:1199 judges each completed compaction using provider-reported prompt usage and records an ineffective attempt when it remains above the trigger.
• agent/context_compressor.py:1305 stops automatic compaction after two ineffective attempts rather than retrying indefinitely.
• tests/agent/test_compaction_anti_thrash.py:130 covers the rough-preflight/real-usage sequence that could otherwise reopen the loop.
• The load-bearing implementation is 7f9485707d0183af0c9f04e1153e1ec9bd98aa68 (fix(compaction): judge the anti-thrash verdict on real usage, not in should_compress), verified as present in current main.