ParseLLMJsonResponse recompiled its ```json fence-extraction regexp on
every call that fell through the direct-unmarshal fast path — i.e. every
fenced LLM response on the graph-extraction path. Hoist it to a
package-level var.
Adds a test (none existed) covering direct, fenced, no-lang and
prose-wrapped inputs, plus a benchmark.
Benchmark, fenced payload (-benchmem):
before: 3205 ns/op, 5712 B/op, 54 allocs/op
after: 1315 ns/op, 693 B/op, 15 allocs/op
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>