TS bulk accessors: getter/setter signature collisions break dual-backend parity (latent)

[rahlk]

PROBLEM

The TS analyzer disambiguates a get x() / set x() pair only in the dict key inside a class's methods map — sig#get / sig#set (TSNeo4jBackend._method_key, cldk/analysis/typescript/neo4j/neo4j_backend.py) — but TSCallable.signature is the same string for both, derived from the property name rather than the accessor kind (e.g. src/models.User.isAdmin). Only accessor_kind distinguishes them.

The two backends then diverge:

• In-memory (cldk/analysis/typescript/codeanalyzer/codeanalyzer.py): _add_class iterates cl.methods.values() and _add_callable does self._callables[c.signature] = c. Both accessors share a signature, so the second silently overwrites the first. Every consumer of _iter_callables() / _callables — get_callables_overview, get_method_bodies, get_decorated_callables, get_callsites_for, get_method — sees one row per pair, last-writer-wins, with the loser's decorators, call sites and body text dropped.
• Neo4j: each accessor is its own :Callable node, so the same query returns two rows sharing one signature, uncollapsed.

Net: overview row counts differ (1 vs 2), bodies and call sites reflect only the surviving accessor in-memory, and decorator handling diverges too — _OVERVIEW_RETURN aggregates with collect(DISTINCT d.name) while TSCallableOverview.from_callable keeps c.decorators with duplicates intact.

SCOPE BOUNDARY

Signature identity for TS accessors, and the decorator dedup semantics that fold into the same family. Not the analyzer's emission — this is an SDK-side identity problem. Not the Python or Java facades.

GOALS

• Fold accessor_kind into TSCallable.signature so a getter/setter pair no longer collides — mirroring the #get/#set disambiguation _method_key already does, but at the signature level so both backends see it consistently.
• Apply the same disambiguation to the in-memory _callables map and _methods_by_class values.
• Reconcile decorator dedup: collect(DISTINCT ...) versus list-with-duplicates, one way or the other.
• Extend the fixture with a real paired get x()/set x() and a callable decorated twice with the same name.

CAVEATS AND KNOWN RISKS

• Latent, not observed. The current sample app has only an unpaired getter (src/models.User.isAdmin), so nothing exercises this today. It surfaces the moment any analyzed app declares a paired accessor.
• tests/analysis/typescript/test_typescript_bulk_parity_live.py (added for #298) runs against that same fixture and therefore cannot see either divergence — the suite is not wrong, it is under-fixtured. Extending the fixture is what makes the parity suite able to catch this.
• Changing signature changes a public, user-visible field. Anyone holding TS signature strings is affected, so this is breaking for accessors and belongs in a major.

DEFINITION OF DONE

• On a fixture containing a paired get/set, both backends return the same row count and the same per-accessor bodies, call sites and decorators.
• A callable decorated twice with one name behaves identically across backends.
• The live parity suite fails before the fix and passes after.