docs(paper): restore reviewed paper in XeLaTeX; omit SQL-DRP from charts

Overleaf was synced to the pre-review paper and its push reverted the revision on
main; restore it. Keep XeLaTeX + xeCJK (Songti TC) for proper CJK glyphs (Overleaf
compiler = XeLaTeX), with all the Codex/Grok review fixes, the e2e-recall column,
the iso-F1 curves, and the F1/F0.5 bar charts.

Represent SQL-DRP honestly in both the paper and docs/recovery-comparison.md: it is
a printable-string carver (space-joined blobs, not per-column records) that scores 0
under the (col1,col2) identity metric -- a capability boundary, not a recovery
failure. Omit it from the score-based charts (paper Fig.2 and recovery-comparison
plot_comparison.py) with the reason in the caption; keep it in the tables with a
footnote. Fix the F1 chart legend overlap.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: h4x0r

docs/img/recovery-comparison.png

@@ -26,14 +26,17 @@
 PNG_PATH = HERE / "img" / "recovery-comparison.png"
 
 # Tool -> colour; category -> marker.
+# SQL-DRP is omitted from the chart: it emits space-joined string blobs, not
+# per-column records, so it scores 0 under the (col1,col2) identity metric -- a
+# capability boundary, not a recovery failure (see recovery-comparison.md). It is
+# kept in the table and prose, but plotting a metric-artifact 0 misrepresents it.
 TOOL_COLOR = {
     "ours": "#1b9e8a",        # teal
     "fqlite": "#e6840f",      # orange
     "undark": "#d62728",      # red
     "bring2lite": "#7570b3",  # purple
-    "sqldrp": "#666666",      # grey
 }
-TOOL_ORDER = ["ours", "fqlite", "undark", "bring2lite", "sqldrp"]
+TOOL_ORDER = ["ours", "fqlite", "undark", "bring2lite"]
 CAT_MARKER = {"0C": "o", "0D": "s", "0E": "^"}
 CAT_ORDER = ["0C", "0D", "0E"]
 

docs/plot_comparison.py

@@ -200,14 +200,18 @@ the same answer key and the same `(col1,col2)` matcher (`BRING2LITE_CMD` =
   [`competitive-landscape.md`](competitive-landscape.md).
 
 ![Precision-recall plane plus F1 and F0.5 grouped bars for ours, undark,
-fqlite, bring2lite, and SQL-DRP across categories 0C/0D/0E](img/recovery-comparison.png)
+fqlite, and bring2lite across categories 0C/0D/0E](img/recovery-comparison.png)
 
 The figure plots the **same harness-computed numbers** as the table above:
 `forensic/tests/nemetz_tool_comparison.rs` writes the per-(tool, category)
 `recall_substrate`, `precision`, `F1`, and `F0.5` to
 [`img/comparison_metrics.csv`](img/comparison_metrics.csv) when run with the
 undark/fqlite oracles, and `docs/plot_comparison.py` renders the chart straight
-from that CSV — chart and table are the same dataset by construction. By **F1**
+from that CSV — chart and table are the same dataset by construction. SQL-DRP is
+omitted from the chart: it scores 0 under the `(col1,col2)` identity metric because
+it emits space-joined string blobs, not per-column records — a measured capability
+boundary (kept in the table and detailed below), not a recovery failure, and a
+metric-artifact 0 would misrepresent it as plotted. By **F1**
 (balanced), sqlite-forensic leads `0C` (0.909 vs fqlite 0.798); under **F0.5**
 (precision-weighted — the forensic β, since a phantom row costs an examiner more
 than a missed low-confidence one) the `0C` lead widens (0.962 vs fqlite 0.798). The `0E`

docs/recovery-comparison.md

@@ -1,14 +1,14 @@
-# Build the SQLite recovery paper (pdfLaTeX + BibTeX).
+# Build the SQLite recovery paper (XeLaTeX + BibTeX).
 PAPER = sqlite-recovery
 
 .PHONY: all clean
 all: $(PAPER).pdf
 
 $(PAPER).pdf: $(PAPER).tex $(PAPER).bib
-	pdflatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
+	xelatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
 	bibtex $(PAPER)
-	pdflatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
-	pdflatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
+	xelatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
+	xelatex -interaction=nonstopmode -halt-on-error $(PAPER).tex
 
 clean:
 	rm -f $(PAPER).aux $(PAPER).bbl $(PAPER).blg $(PAPER).log $(PAPER).out

paper/Makefile

@@ -1,16 +1,17 @@
 % =============================================================================
 % Deleted-Record Recovery from SQLite Databases
 % A reference implementation and a multi-oracle forensic evaluation
-% pdfLaTeX + CJKutf8, two-column  (builds under Overleaf/arXiv default engine)
-% Build: pdflatex -> bibtex -> pdflatex x2   (see Makefile)
+% XeLaTeX + xeCJK, two-column  (Overleaf: set the compiler to XeLaTeX)
+% Build: xelatex -> bibtex -> xelatex x2   (see Makefile)
 % =============================================================================
 \documentclass[10pt,a4paper]{article}
 
 % --- Packages ----------------------------------------------------------------
-\usepackage[utf8]{inputenc}
-\usepackage[T1]{fontenc}
-\usepackage{CJKutf8}
-\usepackage{lmodern}
+\usepackage{fontspec}
+\usepackage{xeCJK}
+\setCJKmainfont{Songti TC}
+\setCJKsansfont{Songti TC}
+\setCJKmonofont{Songti TC}
 \usepackage{amsmath,amssymb,amsfonts}
 \usepackage{booktabs}
 \usepackage{multirow}
@@ -45,8 +46,8 @@
     urlcolor=blue!60!black,
 }
 
-% inline simplified-Chinese helper (pdfLaTeX + CJKutf8)
-\newcommand{\zh}[1]{\begin{CJK}{UTF8}{gbsn}#1\end{CJK}}
+% xeCJK handles CJK natively
+\newcommand{\zh}[1]{#1}
 \newcommand{\orcidlink}[1]{%
   \href{https://orcid.org/#1}{\textcolor{green!50!black}{\textsf{iD}}}%
 }