# Seed corpora

Hand-written starting inputs, one directory per fuzz target. These are
committed; `corpus/` is not.

The split follows `astra-soak-overview`, which calls a minimized corpus
"accumulated compute, not a build artifact". Both halves of that sentence decide
something:

- **These seeds are human intent.** Each one names a shape of the grammar worth
  reaching in the first second rather than the first hour: a filter chain, a
  fenced code block, an unbalanced quote. They are reviewable in a diff, they
  do not churn, and they are what a fresh machine needs to start somewhere
  better than random bytes.
- **`corpus/` is machine output and lives on astra**, under the soak runner's
  persistent directory, where it accumulates across restarts. It is worth
  committing only once it represents real soak hours; a corpus minimized after
  two minutes of local fuzzing is neither accumulated nor compute. Minimize
  with `cargo +nightly fuzz cmin <target>` and commit it when that day comes,
  not before.

Run a target against these on a machine with no corpus:

    cargo +nightly fuzz run substitute seeds/substitute

## Crash seeds

An input that once found a bug stays here forever. Two so far, both from one
root cause: the marker regex is not quote-aware.

- `14-quote-aware-marker-injection`: substitution emits live template syntax.
- `15-close-brace-in-string-arg`: the minimal form, a `}}` inside a quoted
  filter argument truncating the marker.

These are seeds rather than a `tests/regressions/` directory because the crate
is not fixed yet. When it is, they become unit tests too, and they stay here
regardless: a test proves the one case, the seed keeps the fuzzer pushing on
the shape around it.