v0.25.0

v0.25.0 — Honest E7 SWE-bench Verified end-to-end eval

Phases 1–5 (PRs #1–#10 on base d0d497c):
  #1 CI paper build · #2 hbox fixes · #3 real LLMLingua-2 + LongLLMLingua E5 + τ n-expansion
  #4 README/site refresh · #5 n=300 SWE-bench_Lite headline
  #6/#7 v0.24: shuffled-position E5, honest operating-point sweep, LongLLMLingua fix
  #8 E7 SWE-bench Verified end-to-end (headline) · #9 v0.25 PDF refresh
  #10 docs/research.html E7 sync

Headline (E7, PR #8): a real agent (aider + claude-sonnet-4-6) on SWE-bench Verified
(n=50, seed 1729, official swebench harness). Full context 52.0% (26/50); distil at its
certified trunc@500 point only 16.0% (8/50), −36pp, paired McNemar p<0.001; LLMLingua-2
26.0% (13/50). Verdict: the localization decision-equivalence certificate does NOT transfer
to end-to-end task success once compression is aggressive. The certificate (the contract)
is the contribution, not the compressor. Numbers: docs/paper/results/swe_bench_verified_e2e.json.