fix(autograder): grade comment-stripped SQL so guide comments don't skew scores

[lassebenni]

Problem

The Week 10 autograder (.hyf/test.sh) has the same file_is_filled() blanket-"TODO" gate as Week 9 — but Week 10 needs a deeper fix, because its dbt scaffold ships -- TODO: guide comments inside the starter models, and those comments contain gradeable keywords ("add JOIN", "add GROUP BY", the six output column names).

Two failure modes result:

1. Correct work scored as a stub. A student who writes the model below the guide comment and leaves the comment in place was failed by file_is_filled (it rejects any file containing "TODO"). Verified: a fully-correct solution that keeps the comments scores 55/FAIL under the current grader (fct_daily_borough_stats.sql: still contains TODO stubs, singular test: still a TODO stub).

2. Naively relaxing the gate lets the empty scaffold pass. Just fixing file_is_filled (the Week 9 approach) makes the bare scaffold score 62 and pass, because the pattern greps (grep JOIN "$mart") match the JOIN / GROUP BY / column-name keywords sitting inside the TODO comments.

Fix

• file_is_filled() strips SQL line-comments and only flags a line whose content begins with TODO (an unreplaced markdown/YAML placeholder).
• New sqlgrep() helper matches patterns against comment-stripped SQL. Every .sql content check now routes through it: staging source(), IS NOT NULL, fare_amount >= 0, tip_pct; mart ref() to both, JOIN, GROUP BY, output columns; and the singular-test ref(). Keywords hidden in comments no longer count as real SQL. YAML/markdown checks are unchanged.

Verified (both sides)

Case	before → after
Bare scaffold (nothing filled)	62 pass → 48 FAIL ✅
Correct solution, guide comments kept	55 FAIL → 100 pass ✅

So it stops penalizing left-in guide comments and stops the empty scaffold from passing on comment text. There are currently no open student PRs on this repo, so no live submissions are affected; this hardens the template before the cohort forks it.

Note for future weeks: prefer sqlgrep (comment-stripped matching) for any keyword check on a .sql deliverable, and avoid the "fail on any TODO" gate — it both over- and under-counts.

🤖 Generated with Claude Code