Record the post-#906 typed-graph bench: two groups, thresholds read, tokens by phase

[WaylandYang]

Two full groups (100 documents, judge 200, errata) on dev at b9e3aae, with #906 and #916, same corpus, model and judge as the pre-#906 baseline recorded in #917. Written into the bench README as the formal cut-2 measurement.

• Judged precision 81.0% / 79.0% before errata, 83.0% / 80.5% after: over the 75.3% threshold, about ten points under the baseline, and the difference is misworded facts from minimal reasoning on extraction.
• Same-sentence gold recall 14.2% / 14.3% before errata, 17.8% / 17.1% after: under the prototype's 15.3% before errata, over it after, and errata now adds without retracting wrongly (2 and 1 retractions, none judged stated).
• 31k and 34k tokens per document: a fifth of the baseline's 158k / 183k, still five times 0044's ceiling; seven tenths of it in alignment and rule proposals, which are paid per signature and are at their most expensive on a fresh 100-document base. The README says the bench still lacks the marginal-tokens-on-a-warm-base reading that would measure the product's per-document cost.
• 47 and 48 minutes a group, against three to five hours.

🤖 Generated with Claude Code