fix(gateway): make compress responses truthful

[bobashopcashier]

Summary

• make /compress detect when the transcript is still fully protected and return a truthful no-op message instead of reporting fake success
• relabel the success token line as a rough transcript estimate so it is not presented like a real pressure or provider-usage metric
• add focused gateway tests for the no-op and successful compression paths

Related

• Fixes /compress can report a successful compression even when the transcript is unchanged #6202
• Fixes /compress can report higher token counts after successful compaction because the banner uses a rough transcript-only estimate #6217
• Related: fix(gateway): detect /compress no-op and clarify token estimates #6270

Why

• /compress could report success even when no middle segment was actually compressible
• the success response implied the rough transcript heuristic was a real token/pressure measurement
• this PR checks compressibility before calling _compress_context(), so the no-op path does not split sessions, rewrite transcripts, or reset stored token counts

Notes On Related PR

• fix(gateway): detect /compress no-op and clarify token estimates #6270 addresses the same issue area
• this PR keeps the scope similarly small but performs a preflight boundary check from the temporary agent's compressor instead of detecting no-op only after compression returns

Testing

• source .venv/bin/activate && pytest tests/gateway/test_compress_command.py -q

Notes

• left unrelated local worktree changes out of this PR

[bobashopcashier]

Closing to reopen as a refreshed PR that follows the current contributing guide and PR template.

[bobashopcashier]

Superseded by #6944, reopened on a fresh branch with the current contributing-guide/template format.