comment_lines conflates blank lines with real documentation (engine-wide, found via #2593) #2625

Found while triaging #2593 (python's rosetta cross-language-consistency sweep). Not
python-specific — the bug is in the shared comment_lines computation, so it likely biases
every language's comment_lines metric toward languages whose source style uses more blank
lines, not just ones with more actual comments.

The bug

comment_lines is computed in keyword-rosetta's tools/bias_report.py:

struct["comment_lines"] += max(0, (row["total_loc"] or 0) - (row["coding_loc"] or 0))

i.e. total_loc - coding_loc, reading the two DB columns galaxyscope writes per file.

• coding_loc (gitgalaxy/core/prism.py split_streams, ~line 315-322) is computed from
  non-blank lines only: total_active_lines = len([l for l in content.split("\n") if l.strip()]), then coding_loc = len([l for l in final_code.split("\n") if l.strip()]),
  doc_loc = max(0, total_active_lines - coding_loc). Blank lines never enter either side.
• But the DB's total_loc column (galaxyscope.py ~line 640, filter_res.get("total_loc", 0)) is the raw line count from aperture.py's ingestion filter — includes blank lines.

So comment_lines = total_loc(incl. blanks) - coding_loc(excl. blanks) silently folds every
blank line into "documentation," on top of real comment/docstring lines. A file with lots of
blank-line-separated functions reports inflated comment_lines for a reason that has nothing
to do with how well-documented it is.

Evidence (python, keyword-rosetta corpus)

data/python/{main,a,b,c}.py: engine reports comment_lines = 52 total, median across 46
languages is 30 (+73%, red band). Direct count: 31 of those lines are blank, only ~10-12
are real #-comment or docstring content. The blank-line count alone accounts for most of the
deviation — python's shell files happen to have a higher blank-line-to-code ratio than most
other languages' shells (stylistic choice per file, not a signal of anything).

Two independent fixes, either or both

1. Corpus-side (keyword-rosetta): normalize blank-line density across all 46 language
   shells so comment_lines stops being dominated by formatting style. Doesn't fix the
   underlying engine metric, just removes this specific cross-language noise source.
2. Engine-side (this repo): comment_lines/doc_loc shouldn't count blank lines as
   documentation at all. Since coding_loc already excludes blanks, the parallel total_loc
   used for the subtraction should too (or bias_report.py/whatever consumes doc_loc
   directly should use prism.py's own doc_loc output instead of re-deriving via
   total_loc - coding_loc against a blank-inclusive total). Needs a full re-bless per the
   Differential Scan protocol since it changes doc_loc for every language, every file.

Recommend (1) as a quick corpus-side mitigation and (2) as the real fix, scoped as its own
issue since it touches every language's golden master, not just python's.