perf: make Node subclass bytes (C-level hash/equality)

[AmitMY]

Answers "is there a faster counter?" — the bottleneck isn't Counter (it's already C); it's hashing/comparing the candidate node-tuples it's fed. Counter(graph.get_merges()) re-hashes them every step, and merge scans compare them. With Node a frozen dataclass, __hash__/__eq__ are Python calls.

Make a Node be its bytes:

class Node(bytes, GraphVertex):
    __str__ = GraphVertex.__str__   # bytes wins the MRO; keep our str()
    def __new__(cls, value): return super().__new__(cls, value)
    @property
    def value(self): return self[:]
    ...

Now __hash__/__eq__/__len__ are bytes' C-level operations (and tuples of bytes hash ~4× faster than tuples of dataclass instances). It's net simpler (drops the custom __eq__/__hash__/__len__).

Why this one matters for the whole pipeline: it speeds the recount path, so unlike the disconnected-only incremental work (#42) it also helps Boundless and SuperBPE — and the two compose.

Output identical (merge digests unchanged at 50/200 and 100/400), 138 tests pass, memory unchanged (nodes are interned, so the slightly larger object is negligible):

	main	this PR
BPE	0.384s / 1.33MB	0.220s / 1.33MB — 1.7×
BNE n=4	0.768s / 2.63MB	0.337s / 2.63MB — 2.3×
Boundless	0.469s / 1.42MB	0.264s / 1.42MB — 1.8×

Note: Node now compares equal to a raw bytes of the same value (bytes __eq__), but in the graph only Node/NodesSequence are ever compared, so behavior is unchanged in practice.

🤖 Generated with Claude Code