Record two repository facts so the next validation checklist does not repeat them #138

Two of the four #127 boxes that turned out wrong as written failed on facts about this repository that nothing wrote down. Recording them once, where a checklist author will look, stops the next validation issue from asking for the impossible.

The two facts

1. The git history begins at LFS: scope it to sources/**, and stop fetching it in CI #57. This repository has a single root commit, 931d626 (2026-08-10, "LFS: scope it to sources/**"), and 50 commits in total. Everything earlier — the Feb 2025 migration from QuantEcon/data and the 2026-07-16 flatten (Flatten the consumer-keyed tree into the published layout #10) — exists only in prose (manifest headers, PLAN.md). A check phrased as "git log --follow reaches Flatten the consumer-keyed tree into the published layout #10" cannot be satisfied; the blob hash at 931d626 is the earliest thing git can show.
2. check_consumed_files.py hashes every file with a recorded sha256, regardless of consumers. That was the deliberate rekeying in Guardrail: hash a file whenever its manifest records a hash #56 and it is the right behaviour, but it means consumed-file-check goes red for any change to a published file's bytes, consumers or not — measured on Throwaway, do not merge: #127 §3.3 CI-gating check with a deliberate manifest mutation #130. A check that wants the schema gate isolated must mutate a manifest, not the data.

To do

Add both to AGENTS.md, in the section a validation author reads first (the repo map, or beside the existing note that the manifests are the register of record), in two sentences each; if the repo adopts a validation-issue template, the same two lines belong there.

Exit criterion

Merged on main.