@workglow/eval ships a bin and no importable entry point, so its scorer is duplicated downstream

[sroussey]

examples/eval/package.json@0.4.0 has main: null, module: null, types: null, exports: null and only bin: { "workglow-eval": "./dist/workglow-eval.js" }. import { scoreExtraction } from "@workglow/eval" cannot resolve for anyone.

The eleven-cycle observation that "sec imports @workglow/eval zero times" has been read as discipline. It is packaging: there is nothing to import.

The consequence is two independent implementations of one function with two independent result interfaces:

• examples/eval/src/score/extraction.ts:67 (ExtractionScore at :31, 140 LOC)
• @workglow/sec's src/eval/scoreExtraction.ts:354 (ExtractionScore at :46, 505 LOC)

The same non-sharing shows up in the model rate card, and it is now measurably wrong. Of the 10 model ids priced in both examples/eval/src/models.ts and sec/src/config/listPricing.ts, 5 disagree:

id	libs	sec
gpt-5.5 output	15	30
gpt-5.4-mini output	1.6	4.5
deepseek-v4-pro	0.56 / 1.68	0.435 / 0.87
gpt-5.6-terra	2 / 12	2.5 / 15
grok-4.5 cached	0.5	0.3

Last cycle only one disagreed, so this is diverging rather than static.

Asks:

1. add an exports/types map to examples/eval/package.json (~3 lines);
2. retire one of the two scoreExtraction implementations;
3. hoist the pricing data into @workglow/ai and re-export the provider discriminator constants that sec/src/config/registerModels.ts:20-26 says outright are "duplicated here".

Found during the 2026-08-24 review. Snapshot: workglow-dev/prd → analysis/grades/2026-08-24/cross-repo-integration.md.