fix(compaction): skip compression when it cannot reduce the token count

[golldyck]

What does this PR do?

Skips trajectory compaction when it can't actually reduce the token count. compress_trajectory (and its async twin) replaced a compressible middle region with a summary without checking that the summary is smaller than what it replaces. When the region was already small, the "summary" came out larger than the original. So the trajectory grew while still being marked was_compressed=True (observed: 406 tokens went up to 465, tokens_saved = -59).

The fix adds a net-savings guard: when the compressible region's token count is <= the summary target, return the trajectory unchanged instead of expanding it. This applies to both the sync and async paths (sibling call path included).

Related Issue

No separate issue filed; reproduces on current main.

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• trajectory_compressor.py: guard both compress_trajectory and compress_trajectory_async so they skip compression when region_tokens <= summary_target_tokens, returning the trajectory unchanged.
• tests/test_trajectory_compressor.py: sync and async tests asserting a tiny compressible region is left unchanged (no token growth, was_compressed not set).

How to Test

1. On main: compressing a trajectory with a small compressible middle grows the token count yet reports was_compressed=True.
2. Apply this PR.
3. Run scripts/run_tests.sh tests/test_trajectory_compressor.py -q. The added net-savings tests pass (34 passed). Note: 3 tests (*_omits_temperature, kimi/moonshot) fail identically on unmodified main. They're unrelated to this change (temperature-omission for kimi models), not a regression from this PR.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(compaction): ...)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix (no unrelated commits)
• I've run scripts/run_tests.sh tests/test_trajectory_compressor.py; the net-savings tests pass, and the only failures are the 3 pre-existing *_omits_temperature cases that also fail on main
• I've added tests for my changes (sync + async)
• I've tested on my platform: macOS 15

Documentation & Housekeeping

• Documentation: N/A
• cli-config.yaml.example: N/A
• CONTRIBUTING.md / AGENTS.md: N/A
• Cross-platform impact: N/A (pure token-count arithmetic)
• Tool descriptions/schemas: N/A

[tonydwb]

Code Review Summary

Verdict: LGTM

Adds a guard to skip compression when the compressible region is smaller than the summary that would replace it. Prevents token count from growing instead of shrinking.

Looks Good

• Clean guard in both sync and async paths
• Well-tested with both sync and async test cases
• Correctly returns trajectory unchanged when compression would not help

---

Reviewed by Hermes Agent

[golldyck]

Rechecked against current main: the change still rebases cleanly (no conflicts) and the compaction guard tests pass. For transparency, the only red tests in tests/test_trajectory_compressor.py are the three *_omits_temperature (kimi/moonshot) cases, which fail on unmodified main as well — they're unrelated to this net-savings guard. Happy to rebase if you'd like it refreshed before merge.

[teknium1]

Merged via salvage PR #70243 as a clean cherry-pick, authorship preserved — thanks @golldyck! Both sync and async net-savings guards plus your regression tests are on main (merge commit 69339aa). The counterfactual run against pre-fix main confirmed the tests target the real 406→465-token growth bug.