[Kerneta]

What this adds

code-benchmarks/ is a self-contained, offline replication of the .cai vs
Graphify vs raw-files code-retrieval numbers. python code-benchmarks/run_all.py
rebuilds every .cai store from source with the tree-sitter extractor and
regenerates results/RESULTS.md. No API key and no network: structural tools
parse offline and grading is programmatic, so anyone can rebuild every number.

The summary tables now carry two token-reduction columns, .cai vs Graphify
and .cai vs raw, expressed in x (for example 164x less).

Reproduced from source:

• psf/requests (ast oracle): .cai 16/16 at 55 tok, 13x less than Graphify, 164x less than raw
• Real-code retrieval: 10/10 at 24 tok, 5x / 202x less
• Deep multi-hop: 11/11 at 27 tok, 4x / 7x less
• Multi-language (python, js, ts, ruby): 44/44 at 33 tok, 4x / 8x less
• httpx structural, all .cai at 100% recall: imports 218x less, subclasses 73x, tdeps 21x, callers 13x, callees 8x

Benchmark fairness fixes (question sets, not the engine)

• The requests most-imported question is graded any-of, since models and
  compat tie at indegree 10.
• The per-language synthetic set dropped one empty-type question, corrected two
  questions whose gold equalled the query argument, and converted two set
  questions to presence so verbose baselines are not penalised for being
  correct but wordy. .cai still scores full marks.

Engine change

Kerneta-Cai/engine/cai_extract.py becomes a deprecation stub. It is the legacy
AST-only extractor that nothing in the product uses; building a store with it
instead of cai_ts_extract.py silently drops extends edges and under-captures
imports. The stub errors with a pointer to the right extractor, and
Kerneta-Cai/BENCHMARK-REPRODUCE.md documents it so no one repeats the mistake.

Verify

cd code-benchmarks
pip install -r requirements.txt   # tree-sitter grammars
python run_all.py                 # regenerates results/RESULTS.md