fix: reapply compression feasibility guard after model switch

[leavrcn]

Summary

• re-run _check_compression_model_feasibility() after in-place model switches
• prevent context compressors from keeping a threshold recalculated above the auxiliary compression model's real context window
• add a regression test covering the threshold-reset path during switch_model()

Test Plan

• python -m pytest -q tests/run_agent/test_compression_feasibility.py

Notes

This PR intentionally includes only the runtime fix and its regression test. Other local uncommitted changes were left out.

[alt-glitch]

Related to PRs #12316, #10690, #14008, #8786 — all address context-length override resolution across runtime paths. Consider consolidating.

[leavrcn]

Closing — superseded by active work under PR #50384. Will reopen if needed.