fix(neo4j): stream the Cypher snapshot instead of building it as one String

[rahlk]

Problem

--emit neo4j died on ThingsBoard v4.0 (4131 Java files) before writing a byte, and left a zero-byte graph.cypher behind:

OutOfMemoryError: UTF16 String size is 1169246895, should be less than 1073741823
    at CypherWriter.renderCypher(CypherWriter.java:82)

(Reported in the field as Requested string length exceeds VM limit at the same line — same defect, different construction path.)

It looks like renderCypher collected every statement into a List<String> and collapsed it with String.join, and that result has to fit in one array. A script containing any non-Latin-1 character is stored as UTF16, halving the ceiling to ~1.07 billion chars. ThingsBoard renders 1.169 billion. No -Xmx value avoids this.

Fix

CypherWriter.writeCypher(Appendable, rows, appName) streams each statement as it is produced. Nothing larger than one 500-row batch is ever a String, so peak memory is bounded by the batch rather than by the graph. renderCypher remains as a thin wrapper over it — kept for tests and small callers, documented as unsuitable for user-facing output, and delegating so the two cannot drift.

Neo4jEmitter writes through Files.newBufferedWriter in explicit UTF-8. It previously used a bare FileWriter, which encodes in the platform default charset; on Java 11 that is not UTF-8 everywhere, so non-ASCII in identifiers or captured source could already be mangled. Same line, distinct defect — called out rather than folded in silently.

The one subtlety

Statements are separated by newlines, not terminated by them, matching String.join exactly. Getting this wrong appends a single trailing byte — and a test comparing renderCypher to writeCypher cannot catch it, because the former now delegates to the latter and so compares the implementation with itself. I hit exactly this: the first implementation produced 33172816 bytes against the release's 33172815.

So the boundary is pinned directly by statementsAreSeparatedByNewlinesNotTerminatedByThem, and the full script was diffed against the v3.0.0 release build.

Verification

On ThingsBoard v4.0, same command, same input:

	v3.0.0 release jar	this build
exit	1 (OutOfMemoryError)	0
graph.cypher	0 bytes	1.17 GB
wall clock	—	5m20s

The emitted script is well-formed: 4063 UNWIND blocks, all balanced against their ] AS row closers, 20 constraints, correct trailing newline.

On daytrader8 the emitted graph.cypher is byte-identical to the one the v3.0.0 release produces — 33172815 bytes, cmp clean.

Suite: 491 tests, 1 failure, 4 skips — the failure is the pre-existing Docker-dependent CodeAnalyzerIntegrationTest > initializationError, unrelated.

Closes #209