content_hash is never written, so every Neo4j push is a full re-upsert #118

Problem

The Bolt writer's incremental path never works. Every push is a full re-upsert of every module.

bolt.ts reads each module's stored content_hash to decide which modules changed:

const res = await s.run("MATCH (m:TSModule) RETURN m._module AS k, m.content_hash AS h");

but nothing ever writes it. finalizeAnalysis strips the field from the wire envelope
(emit.ts:61, delete mod["content_hash"]), the Neo4j projection is built from that same
envelope, and project.ts:231 says so in a comment before omitting it:

// The wire strips the internal fields (module_name, content_hash) — the graph name is the
// file key, exactly as the historical projection of the stripped tree produced.

So the diff compares against NULL for every module and concludes everything changed, every time.

Confirmed on a live database after pushing superset-frontend:

MATCH (m:TSModule) RETURN count(m) AS modules, count(m.content_hash) AS with_hash;
modules, with_hash
1841, 0

and the log line on a re-push of unchanged source: 1841 modules (1841 changed).

Three parts of the system disagree: schema.ts:97 declares content_hash: "string" on
:TSModule, bolt.ts reads it, and project.ts never writes it.

Why the tests did not catch it

test/neo4j-bolt.test.ts:176 hand-writes the property into a seed node
(CREATE (:Module:TSModule {... content_hash:'stale'})), so the diff logic is exercised against
data the real projection cannot produce.

Parity

codeanalyzer-python does this correctly and is the reference: content_hash is wire payload
(schema/py_schema.py:450, not stripped), projected onto the node (neo4j/project.py:689), and
read back for the diff (neo4j/bolt.py:105). Its incremental path works.

Scope boundary

Make content_hash reach the :TSModule node so the diff functions. Not in scope: the cache's own
content hashing, which is unrelated and works.

Goals

• content_hash is present on every projected :TSModule, matching python
• A re-push of unchanged source reports 0 changed modules, not all of them
• A test that fails if the property stops being written — not one that seeds it by hand
• Decide whether it becomes wire payload in analysis.json (python's choice) or is threaded to
  the projection some other way

Caveats and known risks

• If it becomes wire payload, analysis.json gains a field per module. Additive, and it restores
  parity with python rather than diverging further.
• Enabling a diff that has never actually run may expose latent bugs in the incremental path
  (the per-module purge, edge ownership) that a permanent full re-upsert has been masking.
• Interacts with Neo4j transaction too big #116: with deletion now gated on --eager,
  a working diff means a default push touches only changed modules instead of rewriting everything.

Definition of done

Push a project twice with no source changes; the second run reports 0 changed modules and writes no
node updates, verified against a live Neo4j rather than a hand-seeded fixture.