Fixed environment, fixed inputs, median of three runs; commands, failure cases and corrections included. One question only: does it actually work for this task?
With a random codeword, the baseline still recalled it across sessions — the token was persisted in workspace AGENTS.md and the host's session cache. Cross-session recall is a DSH capability, not a plugin differentiator.
Read the full report →v0.1 did not discriminate; v0.2 moves to post-compaction consistency, unit-level store/recall, and cross-project conflicts.
Read the finding →Environment diagnosis, broken-image triage, batch runs; hit rate, success rate, reproducibility.
Method first →Headless render, glTF integrity, error recovery.
Method first →Publicly installable, version-pinned; validate that the baseline fails before testing candidates.
Record DSH version, model, machine, inputs and time budget.
Three independent runs with fresh HOME and workspace; keep logs and artifacts.
Task-level verdicts, failure cases, reproducible commands; no rankings.