context_compressor: preserve absolute threshold override across model switches

[BeliefanX]

Summary

• Add compression.threshold_tokens as an explicit absolute compression threshold override.
• Preserve that user-configured absolute threshold across model switches, fallback restore, and context-length changes.
• Keep runtime auxiliary compression feasibility lowering separate from the stored user override.

Implementation

• ContextCompressor stores the user/config absolute value in threshold_tokens_override.
• Active threshold is clamped to the current model context without losing the stored override, so switching back to a larger model can restore the intended threshold.
• Auxiliary compression-model feasibility lowering now uses a live-session runtime cap instead of mutating the stored override.
• The runtime cap survives primary-runtime restore/model updates, but does not become user configuration.

Compatibility / review scope

• Existing ratio-based threshold behavior remains compatible when no absolute override is configured.
• This PR stays independent of the gateway/delegation PRs and only covers context-compressor threshold semantics.
• Current GitHub checks are green (success/skipped only).

Testing

• /Users/fanxuxin/.hermes/hermes-agent/venv/bin/python -m pytest -o 'addopts=' tests/agent/test_context_compressor.py tests/run_agent/test_switch_model_context.py tests/run_agent/test_compression_feasibility.py -q — 116 passed
• py_compile for touched runtime/config/test files — passed
• Independent re-review verified explicit override preservation, runtime cap restore, and auxiliary feasibility semantics; no P0/P1/P2 issues.

[BeliefanX]

Refreshed this PR onto current main and resolved the merge conflicts.

Current status after refresh:

• branch is now mergeable against main
• attribution / supply-chain / nix / e2e checks are passing
• the remaining red test check appears to be the current upstream full-suite baseline, not this PR's focused change; failures are broad and match the same unrelated areas seen on other refreshed branches (Bedrock 1M, cron script empty-output note, Telegram topic hint debounce, browser Chromium guard, delegate heartbeat, sandbox cwd assertions, etc.)

I ran the relevant focused tests locally for this PR before force-pushing the refreshed branch.

[teknium1]

Thanks for isolating the requested absolute-threshold behavior and for covering model-switch and fallback restoration.

Problems

• Current main now derives thresholds from the effective input window after reserving model.max_tokens (agent/context_compressor.py:1008-1029). The PR's raw-context_length clamp would bypass that safety contract when salvaged.
• compression.threshold_tokens also needs gateway cache invalidation: gateway/run.py:15623-15634 enumerates baked compressor settings, and the new key is absent.
• The added user-facing setting needs docs. Current documentation states that threshold_tokens is always derived from threshold × context_length (website/docs/developer-guide/context-compression-and-caching.md:167-175).

Suggested changes

• Integrate the override into the current _compute_threshold_tokens()/small-context safety model rather than restoring the older recomputation path.
• Add the gateway cache key and a regression test, then update the compression configuration documentation.

Automated hermes-sweeper review.

[BeliefanX]

Closing this version after re-review against current main. The permanent absolute-threshold knob is not a confirmed product requirement, and this branch's raw-context path no longer fits the effective-input/output-reservation safety model. If a concrete need appears, it should return as a smaller current-main change with the required cache invalidation and documentation.