Fix end-edge attribution (off-by-one), node-id collisions, and stars

[deerdear]

Hi! I'm using this dataset for a visualization project — thank you for publishing it. While validating it against two independent transcriptions of the book, I found a systematic extraction bug and some smaller ones. This PR fixes them.

The main bug: end edges are attributed one pattern forward

In addEdges, lines[i - 2] is the paragraph before pattern N's heading — which is pattern N−1's closing links — but the code attributes those references to N:

for string in re.findall(r'\((\d+)\)', lines[i - 2]):
    e = g.add_edge(node, nodesDict[int(string)], True)   # node = N, should be N-1
    e['origin'] = 'end'

Evidence (validated against the apl-md hand transcription and the iwritewordsgood.com mirror):

• As attributed, only 141/1053 end edges match the book's link structure. Re-attributed to the previous pattern, 1008/1053 (96%) match.
• It also explains all 66 self-edges in the file (pattern N−1's closing referencing N becomes N→N).
• Whole-graph agreement (undirected Jaccard vs the apl-md transcription): 0.34 → 0.81 after the fix.

The fix tracks the previous heading's node and attributes lines[i - 2] to it, plus a guard against self-references and duplicate edges (the book declares most links in both patterns' texts).

Smaller fixes

• "A ROOM OF ONE's OWN" (lowercase s) — pattern 141's heading never matched, silently dropping its links and stars.
• Includes the three fixes from your unmerged fix-data-errors branch (merged in, credited to that commit): (73, 'THE FAMILY') → 75, (201, 'THICKENING THE OUTER WALLS') → 211, and the range(26, 233) subsection typo. The id collisions also silently misrouted both patterns' edges and stars via the nodesDict overwrite.
• .gitignore for __pycache__ (the branch had committed bytecode).

The committed patterns.graphml

I don't have a .txt of the book, so I couldn't re-run the pipeline. Instead the committed file was mechanically repaired with an equivalent transformation: rename the duplicate nodes (73→75, 201→211 + section/subsection), shift every origin=end edge's source back by one, drop self-edges/duplicates, and correct 64 stars values cross-checked against the book mirror (231/253 headings parseable; apl-md confidence tags as fallback). Edges for the four patterns affected by the id collisions were unrecoverable from the file (both patterns' links landed on one node id) and were restored from the apl-md transcription's link metadata.

Result: 253 nodes with unique ids 1–253, 1,474 directed edges, zero self-edges/duplicates, Jaccard 0.81 vs apl-md.

Known remaining limitation (out of scope)

Even after the fix, edge recall is ~60% — the source text conversion missed a large share of the book's cross-references (the book has ~1,750+ links). That's a property of the input .txt, not the script; I can share my comparison tooling if useful.

Happy to split this differently (e.g. rebase onto fix-data-errors) if you'd prefer.