Graph collapses every comment into one docstring string: file-level comments, non-javadoc comments, comment spans and body-node comments are all dropped #257

Problem

The second half of the JSON-vs-graph sweep that produced #255 and #256. Comments are separated out because closing this gap is not a property addition — it needs a decision about what carries a comment on the graph.

analysis.json models a comment as {content, span, is_javadoc} and attaches lists of them to the module, to types, to callables, to fields, to enum constants, to record components, and to individual body nodes. The graph collapses all of that into one docstring string per declaration, built by joining the javadoc entries (V2GraphProjector.docstringOf, line 699). Four distinct facts are lost. Counts from daytrader8 at -a 1:

• File-level comments vanish entirely. 465 comments hang off modules in the fixture; :JModule has no comment property at all, so the license header, the file preamble and every free-floating comment between declarations are simply not in the graph.
• Non-javadoc comments attached to declarations vanish. 398 comments are attached to types, callables and fields; 328 are javadoc and survive as docstring; the remaining 70 are dropped because docstringOf filters on isJavadoc().
• Every comment's position is lost. The span is discarded when the list joins into a string, so no comment can be located in the file or sliced out of :JModule.source, which the graph now carries.
• bodyNode.comment is dropped. 140 of the 4027 body nodes carry one.

The catalog documents the collapse as intentional ("javadoc collapses to a docstring property", V2SchemaCatalog:33), and python-sdk documents living with it: get_all_comments and get_comment_in_file raise rather than answer with a smaller set claiming to be every comment (cldk/analysis/java/neo4j/neo4j_backend.py:92-100). That is an honest workaround for a projection that cannot answer the question, not evidence the projection is right.

The invariant being restored: the JSON payload and the graph carry the same facts. docstring is a legitimate convenience derivation; it is not a substitute for the comments it was derived from.

Scope boundary

The comment model on the Neo4j projection. docstring stays, whatever shape this takes — consumers use it and it costs nothing to keep as a derived property.

Does NOT touch the JSON projection, which is already correct. Does NOT move schema_version, held at 2.0.0 per codellm-devkit/.github#50.

Goals

• decide the shape: a :JComment node family reached by a containment edge from whatever it is attached to, versus a serialized-JSON property per declaration (the parameters_json precedent), versus parallel comments / comment_spans list properties
• file-level comments are reachable from :JModule
• non-javadoc comments are carried, distinguishable from javadoc ones
• each comment's span is carried, so it can be located in the file and sliced out of :JModule.source
• bodyNode.comment is carried
• docstring is retained on every declaration that has it today
• V2SchemaCatalog declares the shape, schema.neo4j.json is regenerated, and the conformance test asserts comment counts per node against the JSON rather than by presence

Caveats

• This is a contract-shape decision, not a property addition. If the answer is a new node family, it adds a label and an edge type to a shared graph contract, which is schema v2 evolution and belongs in design mode rather than in-place work. Triage it before implementing.
• Whatever shape is chosen, codeanalyzer-python and codeanalyzer-typescript inherit it. Both are worse off than this analyzer today: python projects a docstring per class/callable/attribute and drops module comments (codeanalyzer/neo4j/project.py:783,806,825,854), and codeanalyzer-typescript projects no comment or docstring property anywhere at all. A shape decided here is the shape all three must spell the same way.
• schema_version does not move, so the addition is not detectable from the payload. Detection is by presence.
• The cheap-looking option is the lossy one. Joining comments into a string is exactly what produced this defect; a shape that keeps content and span together is the point.

Definition of done

• Every comment in the fixture's analysis.json is reachable on the graph, counted per module and per declaration, with zero shortfall.
• Each comment's carried span slices :JModule.source to that comment's content, byte-for-byte — the check that proves the position survived rather than merely being declared.
• docstring values are unchanged from before the change on every node that had one.
• python-sdk's reason for raising in get_all_comments no longer holds, stated explicitly in the closing note so the SDK-side uptake can be filed against it.
• Emitted schema matches the checked-in snapshot; existing suite green.