Review vendor coverage

The principle (per CLAUDE.md): anything that decides a score is vendored verbatim from NeMo-Skills. SE code is allowed only for transport, glue, default config, and CLI / output formatting.

This skill audits whether SE code has leaked scoring logic. It does NOT auto-fix; vendoring more is a design call the user makes after seeing the report.

1. Enumerate SE files

find sgl_eval scripts tests -name "*.py" \
  | grep -v _vendored | grep -v __pycache__ | sort

2. Classify each into one of four buckets

For each file (and each function within larger files), ask: "Does this code decide a score?"

Bucket	Description	Examples	Action
Transport	Wraps an external service or a Python primitive. No scoring decisions.	sampler, runner, cli, types, registry, metrics formatter	OK as SE.
Glue	Bridges between vendored API and our runtime model. Necessary because of architectural difference (sync vs async, in-memory vs file, single vs batch).	_predictions.py, _loader.py, _eval_single_sync, _score_via_eval_mcq	OK as SE. Document the upstream API it bridges.
Default config	Per-benchmark sampling / repeat / thinking defaults. Model-dependent, not a scoring decision.	_TABLE rows in _registry.py	OK as SE. Confirm tracks NS-aligned defaults.
Score-deciding	Code that parses generation, compares answers, computes pass@k, renders prompts beyond str.format.	(should be empty in SE!)	RED FLAG. See step 4.

3. Specific spots to scrutinize (grey areas)

Known borderline pieces -- not violations, but watch for drift:

Location	What it does	Why grey
evals/_prompts.py:render_prompt	yaml.safe_load + str.format on vendored yaml	Re-implements upstream's prompt subsystem (~500 lines). Equivalent today because templates only use {problem} / {examples}. Drift if upstream adds richer templating.
evals/_math.py:aggregate_with_math_metrics (padding)	while len < n_repeats: append dup	NS doesn't pad; we pad to keep MathMetrics happy under partial sample failure. Affects score only when sampler fails partway.
evals/_math.py:_flatten_math_metrics and _multichoice.py:_flatten	Picks fields out of MathMetrics.get_metrics() nested dict	Display selection, not computation. Safe if upstream keeps the same field names.
_TABLE thinking / n_repeats	Per-benchmark sgl-eval defaults	Not a violation; runtime config. CLI override always wins.

If a new grey area surfaces during the audit, add a row.

4. If you find score-deciding code in SE

Don't auto-edit. Report to the user:

1. What it computes (the scoring decision in plain words).
2. Where in NS upstream the equivalent lives (or "not in NS").
3. What it would take to vendor it: file path(s), import rewrites, any drop_imports/drop_functions needed.

Then ask the user whether to:

• vendor (run vendor-update workflow with extended SOURCES.yaml),
• keep as SE (document the deviation, e.g., in CLAUDE.md grey-area list),
• or remove the logic entirely.

5. Output format

Write a short report (under 300 words):

SE files: <count> total
  transport       : <n> (<list>)
  glue            : <n> (<list>)
  default config  : <n> (<list>)
  score-deciding  : <n> (RED if > 0)

Grey areas reviewed:
  - <location>: <still acceptable / has drifted because X>

New findings:
  - <none> | <list>

Recommendation: <one sentence>