feat: add 5 new agentic benchmark scenarios

[jannemakela]

Summary

Adds 5 new agentic test scenarios that cover real-world in-house legal workflows typically handled by corporate legal teams. All scenarios are multi-turn and require read-reason-edit patterns.

New Scenarios

ID	Name	What it tests
comment-driven-edit	Comment-Driven Edit	Read margin comment, verify change, reply to comment
multi-location-update	Multi-Location Update	Rebrand company name across body, headers, footers, signature block
defined-term-insertion	Defined Term Insertion	Add defined term to definitions + propagate to another clause
liability-cap-rewrite	Liability Cap Rewrite	Read numeric value, reason about format, update word + numeral
clause-deletion-and-renumber	Clause Deletion & Renumber	Delete section entirely + renumber subsequent sections

Why These Scenarios

These represent the most common agentic editing patterns in contract review workflows:

• Comment interaction (comment-driven-edit) — the core review loop
• Global consistency (multi-location-update) — name/term changes must be exhaustive
• Multi-step reasoning (defined-term-insertion) — add + propagate, requires understanding document structure
• Read-compute-write (liability-cap-rewrite) — must parse both textual and numeric representations
• Structural modification (clause-deletion-and-renumber) — must understand section numbering

Changes

• src/scenarios.ts — 5 new scenario definitions (all isAgentic: true)
• src/success.ts — success check functions for each new scenario
• src/scenarios.test.ts — updated to expect 15 scenarios with assertions
• README.md — updated scenario list

Tests

All 11 tests passing ✅ (lint + format + vitest)

---

Contributed under AGPL-3.0-only per repository license.

[jannemakela]

Closing — scenarios need to be rebuilt with real open-source legal documents (Common Paper, NVCA, etc.) and realistic end-to-end workflows (form fill, contract clone, policy check, playbook commenting, multi-file deals).