The symbol indexer ingests __corpus__ byte fixtures the differential explicitly excludes

[nedonatelli]

What

src/agent/tools/__corpus__/ holds deliberately malformed byte fixtures — BOM
variants, CR/CRLF/mixed line endings, NBSP, trailing whitespace, no-final-newline
— used to test file-handling edge cases. They are not source.

The symbol indexer indexes them anyway: 13 files contributing 21 symbols
(17 variable, 4 function) to the live graph.

src/agent/tools/__corpus__/blank-lines.ts   {'variable': 2}
src/agent/tools/__corpus__/bom-crlf.ts      {'variable': 2}
src/agent/tools/__corpus__/cr.ts            {'function': 1}
src/agent/tools/__corpus__/mixed-eol.ts     {'variable': 3}
...

Why it matters more than 21 symbols suggests

scripts/graph-differential.mjs already names this directory in
EXPECTED_MISSES:

const EXPECTED_MISSES = [
  /\/__corpus__\//,   // deliberately malformed byte fixtures — not source
  /\.(md|json|txt)$/, // not code
];

So two components disagree about what counts as source: the differential treats
__corpus__ as legitimately un-indexable, while the indexer ingests it. The
check meant to detect extraction gaps is calibrated against a different file set
than the thing it checks — small here, but it is the differential's whole job to
be the independent view.

These symbols also compete at retrieval like any other, so asking about
line-ending handling can surface a fixture named cr.ts over the code that
handles it.

What to do

Exclude __corpus__ from indexing, ideally by making the indexer and the
differential read the same list rather than maintaining the judgement in two
places — the same convergence INDEX_EXCLUDE_DIRS did for the four scanners.

Worth checking for sibling fixture directories (__fixtures__, __snapshots__,
testdata/) while in there; the reason to exclude them is identical and only
__corpus__ happens to have been noticed.

How it was found

Reconciling the live symbol graph against a direct tree-sitter run after the
tree-sitter fixes (#47/#50/#51) landed. Every per-file count matched exactly;
the only aggregate differences were these 13 files and one file over the size
cap. No defect in extraction — the file sets differed.