bug(unic-archon-dlc): the design contract's reads lost four findings between two runs at identical measurement density

What the two runs' contracts show

Same component, same command version, six days apart. Measurement density is identical — 28 node ids, 15 hex values, 4 pixel figures, 92 bullets in both card contracts. Run 2 lost no facts.

It lost four findings.

Run 1 scoped its four to "inside the has badge cells" — it read into the nested instance layers. Run 2 recorded those same cells and did not read into them. All four have since been filed as design questions in the Consumer's tracker.

Both contracts name the same reads in their prose, and figma.md does list get_design_context. So this is not a missing entry in the read table. It is a missing depth.

The uncomfortable half

The one case where run 2 explains itself is the animation annotation. Its contract says:

"Read 2026-08-26 and recorded in docs/agents/figma.md; not re-measured on 2026-08-31. None of this run's reads surfaced an annotation."

The document existing is why the observation was not repeated. That is the cost side of writing facts down, and it is the only self-explained instance — the other four are unexplained.

The open question

Should the contract's read list require a nested-layer pass explicitly, and how does a run know it has read deeply enough rather than merely read the named surfaces?

And separately: when a fact is already recorded in a Consumer document, should a run re-measure it or cite it? Run 2 cited, said so honestly, and lost the measurement.

Full evidence: docs/research/2026-08-31-dlc-run-2-findings.md.

Start with docs/research/2026-08-31-dlc-run-2-findings.md and compare its evidence with docs/agents/figma.md, especially the get_design_context read entry. Trace how the contract handles nested instance layers and previously recorded facts. Done means the required read depth and the re-measure-versus-cite behavior are explicitly resolved and documented.