scan-all pollutes characterization fixtures (tests/fixtures/char_graphbuffer) with README_AI

[dreamlx]

Problem

codeindex scan-all at the repo root uses the root .codeindex.yaml, whose
include: lists tests/. That makes scan-all walk test fixtures —
including tests/fixtures/char_graphbuffer/, the #101 characterization fixture
that must stay byte-identical — and generate README_AI.md inside them.

Those generated files are not gitignored (README_AI isn't in .gitignore),
so a careless git add -A commits them. That drifted the #101 baseline goldens
and broke test_structural_readmes /
test_enrich_prompts_and_frozen_ai_readmes on CI in PR #134.

Why cb803b8 didn't prevent it

cb803b8 force-added tests/fixtures/char_graphbuffer/project/.codeindex.yaml
so the characterization test's own scan (copytree + stub ai_command) behaves
identically locally and in CI. But that config only governs the project/
subtree when the test runs it — it does not stop the repo-root scan-all
from indexing char_graphbuffer/ from the outside.

Repro

```bash
codeindex scan-all
git status --short | grep char_graphbuffer # README_AI.md generated at 4 levels
```

Fix options (need a decision)

1. Exclude characterization fixtures from the root scan — but the root
   .codeindex.yaml is gitignored (per cb803b8), so a local exclude won't
   reach CI. Would need a tracked exclude mechanism.
2. gitignore README_AI under fixture dirs — e.g. add
   tests/fixtures/**/README_AI.md to .gitignore so generated fixture READMEs
   can never be committed. Lowest-risk; the fixtures' own READMEs (where they
   exist as legit test assets) would need ! un-ignore exceptions.
3. Mark the fixture as scan-all-immune via a sentinel (empty .noindex or
   similar) that scan-all respects.

Current workaround (applied in #134): manually delete the polluted READMEs
before committing. Not durable — the next scan-all + git add -A re-breaks.

Priority

Low-maintenance landmine: only fires when someone runs scan-all and commits
the result. But it silently breaks CI with a confusing characterization-drift
failure, so worth a real fix.

Discovered while dogfooding codeindex + loomgraph (#132/#133 PR).

[dreamlx]

Residual found & fixed in #170: #136's gitignore stopped the pollution from being committed but not from sitting on local disk — copytree carried the gitignored (locally invisible) README_AI.md files into the characterization scan, and GH #38's enrichment-preserve logic kept their stale <!-- enrichment: ok --> markers → golden drift. CI stayed green (clean checkout) while local runs failed permanently. #170 makes the fixture copy ignore README_AI.md.