Grounding fingerprints are hex-encoded JSON and take up to 39% of a scaffold

[theyashasvipandey]

Grounding fingerprints are committed to Markdown as hex-encoded JSON, and in a
setup-populated scaffold they take up more than a third of the scaffold's bytes. Every agent that
reads a context file pays for them in tokens.

Measured

Share of .mex/**/*.md bytes on lines that are fingerprint: or bodyHash::

scaffold	fingerprint/hash share
project set up with 0.8.2	39.3%
Hono, populated by setup	37.4%
two older populated projects	27.0%, 26.3%
mex itself	14.2%

The average fingerprint across 66 entries is 1,704 characters.

Cause

serializeFingerprint() (src/graph/fingerprint.ts:81):

const payload = Buffer.from(JSON.stringify(fingerprint), "utf8").toString("hex");
return `${FINGERPRINT_PREFIX}:${K}:${payload}`;

JSON of 64 decimal uint32 values plus full neighbour ids (function:<32 hex>, 5.3 on average),
then hex-encoded, which doubles it.

Options (sizes are averages over the same 66 entries)

encoding	avg chars	risk
today: hex(JSON)	1,704	—
A. plain JSON, or base64url(JSON)	~850 / ~1,140	none; same data
B. base64url of the 64 minhash values as binary, plus neighbour ids shortened	~400	neighbour shortening changes what reconciliation compares; needs a collision check
C. move fingerprints to a committed sidecar file, keeping only node + bodyHash in frontmatter	~2–3% of scaffold	new file format; needs a design

Whatever the encoding, use a new prefix (for example mh2:64:) so deserializeFingerprint
reads both and existing scaffolds keep working. Rewrite old entries only when a file's groundings
are already being rewritten.