fix(accuracy): OIA auto-classifier produces near-universal false positives from .github boilerplate

The heuristic classifier (scripts/classify-oia.mjs) builds a single corpus from the repo file tree + README + topics + language, then regex-matches signal groups. Because raw file paths are in the corpus, GitHub boilerplate that exists on most repos drives two of the most "reliable" signals.

Reproduced on sindresorhus/slugify (a trivial string utility):

• classifies as L7 Orchestration & Workflow (presence) because .github/workflows/ and .github/workflows/main.yml match /workflow/
• gets a security span because .github/security.md matches /security/

Secondary: several groups lack word boundaries, so common English inflates hits (e.g. /knowledge/ is a substring of "acknowledged").

Impact: L7 and the security span are close to noise across the matrix; self-service submitters can also salt a README to place themselves on flattering layers.

Proposed fix (see companion patch):

1. Exclude boilerplate from the corpus before matching: .github/, node_modules, lockfiles, LICENSE, SECURITY.md, CODE_OF_CONDUCT.
2. Add \b boundaries to substring-prone tokens (knowledge, train, serve, vector).
3. Weight README + topics above raw tree paths rather than a flat equal-footing concatenation.

Must be applied to BOTH the server classifier and the retest() twin in docs/oia-matrix.html (see the twin-sync issue).

Start in scripts/classify-oia.mjs by tracing how the corpus is assembled and how signal groups are matched, then compare the retest() twin in docs/oia-matrix.html. Verify the boilerplate exclusions, word boundaries, and README/topics weighting against the slugify reproduction, and ensure both classifiers stay synchronized.