Compact routed repository guidance preserved correctness while reducing coding-agent tokens

General lesson: for bounded repository coding tasks, a short always-loaded safety/bootstrap file plus a task router can preserve or improve correctness while reducing token use; keep deep domain documentation discoverable, but do not force agents to read history, plans, audits, vendor references, bug registers, or unrelated skills at startup. This applies to repositories with substantial Markdown guidance and deterministic task checks. It does not establish that domain guidance is unnecessary for architecture, incident response, live integrations, or other high-context work.

Evidence (2026-09-19): In a blinded Shield experiment using GPT-5.6-Sol medium, 108 scored BB sessions covered 12 deterministic repair tasks. Current documentation passed 46/48; compact routed documentation passed 48/48; a safety-only control passed 12/12 with one repetition per task. Compact routing used 17.0% fewer total tokens, 9.9% fewer uncached input tokens, and 9.0% fewer output tokens than current docs. Task-bootstrap 95% intervals excluded zero for those three token measures. Mean elapsed time was 8.9% lower, but its interval included zero. Two failures both occurred with current docs and were semantic target-selection mistakes. There were no harness-invalid scored trials or final protocol violations. Source record: https://linear.app/coreedgesolution/issue/COR-4353/evaluate-shield-repository-context-with-expanded-coding-agent

Inference: the useful intervention is routing and removing mandatory reads, not deleting canonical docs. Uncertainty: the safety-only condition was smaller, and all tasks were bounded offline repairs on one repository and one model.