fix(writeup): sanitize abstract before render_tex

The abstract string on the Paper dataclass comes straight from
idea.get("Abstract") (produced by the ideation step), not from the
per-section LLM pass, so it bypassed the SANITIZE_PIPELINE entirely.
Real-world demo run hit "Missing $ inserted" on the very first
pdflatex pass because ideas.json contained "perplexity_gap" — an
underscore LaTeX tries to open a subscript with.

Running _sanitize_latex on the abstract covers the same failure modes
as every other text block (underscore escape, orphan $, prose
specials, reasoning-tag leakage, etc.) for ~zero cost.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: easyvibecoding

skills/hermes-sci/package/hermes_sci/writeup.py

@@ -413,9 +413,16 @@ def context_fn(section_key: str) -> str:
                 log.info("inserted table labels in %s: %s", k, sorted(after - before))
         cleaned[k] = v2
 
+    # Abstract comes from idea metadata (ideation step), not from the
+    # per-section LLM pass, so it also bypasses the per-section sanitize
+    # pipeline. Run it through explicitly so prose specials like `_` or a
+    # truncated inline equation don't crash pdflatex before \section{} even
+    # starts.
+    abstract = _sanitize_latex(str(idea.get("Abstract") or ""))
+
     return Paper(
         title=str(idea.get("Title") or "Untitled Research"),
-        abstract=str(idea.get("Abstract") or ""),
+        abstract=abstract,
         sections=cleaned,
     )
 

skills/hermes-sci/package/tests/test_math_balance.py

@@ -61,6 +61,22 @@ def test_triple_dollar_in_prose_drops_last():
     assert out.count("$") == 2  # one removed → even count restored
 
 
+def test_abstract_with_underscore_sanitized_by_writeup():
+    """Regression: ideas.json produces abstract text that bypasses the
+    per-section sanitize pipeline (it comes from ideation, not writeup).
+    write_paper must sanitize it explicitly before render_tex, otherwise
+    `perplexity_gap` etc. crashes pdflatex on page 1."""
+    from hermes_sci.sanitize import sanitize_latex
+    abstract = (
+        "We train on (input, perplexity_gap) pairs where perplexity_gap "
+        "measures draft-target discrepancy."
+    )
+    out = sanitize_latex(abstract)
+    # Both occurrences escaped.
+    assert out.count(r"perplexity\_gap") == 2
+    assert "perplexity_gap" not in out.replace(r"\_", "@")
+
+
 def test_dollar_inside_display_math_does_not_mask_prose_orphan():
     """A bare `$` in prose + a self-contained display-math block
     should still detect the prose `$` as orphan."""