sem_edges reads wrong key from graph.json

[Joseph94m]

Problem

The edge merge logic in watch.py _rebuild_code() reads from existing.get("edges", []) but graph.json uses the key "links" (NetworkX node-link format). This means sem_edges is always an empty list and all semantic edges are silently dropped on every watcher rebuild.
This is the root cause behind #253 and #261. The fixes in 0.4.3, 0.4.4, and 0.4.5 added correct merge logic but all read from the wrong key, so the edge preservation never actually executes.
Node preservation works (reads "nodes", correct key). Hyperedge preservation works (reads "hyperedges", correct key). Only edge preservation is broken.

Fix

One line in watch.py, line 46:

# Before (broken)
sem_edges = [e for e in existing.get("edges", [])
# After (working)
sem_edges = [e for e in existing.get("links", existing.get("edges", []))

Falls back to "edges" for safety in case a future version changes the key.

Verification

Tested locally by patching the installed 0.4.5 watch.py:

Metric	Baseline	Unpatched watcher	Patched watcher
Total nodes	587	588	588
Total links	534	396	531
INFERRED	49	13	47
AMBIGUOUS	2	0	2
Edges touching docs	138	3	138
Doc INFERRED	35	1	35
Hyperedges	22	22	22
Patched watcher preserves nearly all semantic edges. The small delta (534 vs 531 links, 49 vs 47 INFERRED) is from edge deduplication during rebuild, not data loss.

[safishamsi]

Fixed in 0.4.7 — now reads . Semantic edges are fully preserved across watcher rebuilds.