## Round 20 update — B + A applied (v3 runs)

Flags: CLI default (decompose ON, max_seq_length 4), cap raised to 1e30.

| codebase | grammars | pure bags | capped scores |
|----------|----------|-----------|---------------|
| RAGSAK   | 95 (was 29)  | 5 (was 9)   | 0 (was 20@1e12) |
| fastapi  | 118 (was 109)| 26 (was 74) | 0 (was 83@1e12) |
| zod     | 10 (was 16)  | 5 (was 15)  | 0 (was 15@1e12) |

- B (decompose default): more grammars, far fewer pure bags.
- A (cap fix): lang_size_score no longer saturates; bags vs tight now rank
  correctly (tight=20, bag=9975 in unit check). mdl_score still shows 0 for
  many small grammars (separate model_cost/lang_size display quirk, not the
  langsize scorer).
- Quality is two-tier: ~10-15% are real sequential flows
  (e.g. web controller test: post->jsonPath->isEqualTo->exchange->expectStatus);
  ~85% remain orderless bags, concentrated in LARGE groups (tests, v4/locales).
- Bags survive because CRX emits one grammar deterministically; lang_size_score
  only ranks BETWEEN algorithms. iDRegEx (off by default) would let the now-fixed
  scorer substitute tighter grammars. Next: run iDRegEx only on bag groups.