A field test is an empirical proof that the implementation was reproducible
from outside at one moment in time. Once that proof is recorded, the corpus is
spent — so nothing durable may cite it: a pointer into it either rots when the
corpus is cleared, or keeps a dead surface on the agenda of every agent that
greps the tree.
Five sites cited it. Each is replaced by the description it was standing in
for, never by a second pointer:
- `aura-cli/src/diag.rs` named a captured `.err` file for the pluralization
bug. The sentence beside it already carried the finding; only the path went.
- `aura-engine/src/construction.rs` opened the acyclicity test with "Mirror
<corpus path>", then described the two-cycle topology in full immediately
below. The description was already adequate; the pointer was redundant.
- `c24-blueprint-data.md` cited a corpus blueprint as the worked example for
the serialized value forms. It now cites `crates/aura-cli/examples/r_sma.json`
— a shipped example carrying both an `{"I64":…}` and an `{"F64":…}` bind,
which is what the passage is about.
- `authoring-guide.md` claimed its inline blueprint was byte-identical to a
corpus file. The guide shows the document in full; the equality claim served
nothing and was dropped for the property that matters — every command in the
guide runs against exactly that document.
- `c24-blueprint-data.history.md` carried a corpus path inside an archived
milestone note. The historical claim (a green end-to-end field test proved
the story) is untouched; only the pointer is gone.
`.gitignore`'s `/fieldtests/*/target/` stays: it points at nothing, it keeps
4.2 GB of build artifacts out of the repository.
Suite 1652 passed / 0 failed (exit 0), including
`guide_worked_example_matches_the_scaffold_template`, the one test that reads a
file changed here.
claude
a3cd6a5b68