feat(v0.4-alpha): llmlingua_style selector + 11-family paired-task corpus

[snchimata]

Summary

Advances v0.4-alpha (Phase 8, F-057) toward its "reproducible baseline report" exit: adds the last buildable deterministic selector and expands the paired-task corpus from 3 to 11 families. Still shadow-only, no ML, nothing served — develop-targeted research.

What's in it

• llmlingua_style selector — a perplexity-free LLMLingua-style proxy: ranks units by mean per-token self-information (surprisal, -log2 P(token)) under a document-derived unigram model. A deterministic stand-in for LLMLingua's small-LM perplexity (the real method needs an LM at inference, deferred per model-research.md). That's the last lightweight selector; RTK (external tool) and the Kompress-v2 sweep (ML checkpoint) remain deferred.
• Corpus 3 → 11 fixtures covering the model-research.md required slices: long_context_needle, code_build_error, code_patch, tool_call_json, log_multi_service, ccr_marker, and the project-disjoint rust_holdout + typescript_holdout hard slices.
  • Authored + self-validated against the live gate in isolated temp dirs, then adversarially verified (8/8 confirmed). Each passes the gate and genuinely differentiates selectors (≥1 selector fails the task at the 25% ceiling), so the report discriminates instead of trivially scoring 1.0.

The report now discriminates

baseline                 ratio   task   crit  achieved
keep_all                  0.25    1.0    1.0       1.0
bm25                      0.25 0.8182    1.0    0.2437   ← strongest deterministic selector
frequency                 0.25 0.2727    1.0    0.2437
llmlingua_style           0.25 0.2727    1.0    0.2448
recency                   0.25    0.0    1.0    0.2422
forced_only               0.25    0.0    1.0    0.1173   ← safety floor
deterministic-tokenfold   0.25    1.0    1.0     0.937   ← safe, but can't hit the budget

Two gaps a learned selector must close: match bm25's 0.82 task success at aggressive ratios (and be more robust across task types), or match tokenfold's safety (1.0) at real compression.

Testing

python eval/run_baselines.py --gate → pass, 132 checks (6 selectors × 11 fixtures × 2 ratios): universal critical-atom survival, ceiling respected, keep_all upper bound.

Still deferred

RTK / RTK+tokenfold (external tool) and the unmodified Kompress-v2 achieved-token sweep (ML checkpoint); Tier-B public-repo corpora with license/revision manifests + project-disjoint splits; structural segmentation; LLM-judge failure diagnosis.