[EPIC] tdbin adoption blockers for a real multi-language consumer (Rust + TypeScript + Kotlin)

[MelbourneDeveloper]

Tracking issue for everything typeDiagram needs to ship before a multi-language consumer can actually adopt tdbin.

The consumer

Deslop is a live duplicate-code analysis engine with four surfaces generated from one .td schema (docs/models/live-ipc.td, ~50 types):

Surface	Language	Wire types today
deslop-lsp	Rust	generated (--to rust + post-processing)
deslop-mcp	Rust	generated (same module)
VS Code extension + webview	TypeScript	generated (--to typescript + post-processing)
JetBrains plugin	Kotlin	hand-written — no target exists

Everything currently moves as newline-delimited JSON. One live report over our own repo is 797 KB / 712 clusters / 1688 occurrences across 475 files, re-serialised and re-parsed on every incremental generation, then again on every editor query. This is precisely tdbin's use case.

Note on scope: LSP and MCP both mandate JSON on their public transports, so tdbin would carry the private deslop-lsp ↔ deslop-mcp IPC channel and the on-disk report/cache artifacts — the two places that dominate our serialisation cost and are entirely ours to choose.

Blockers

• [BLOCKER] Publish the tdbin Rust runtime crate — generated codecs reference a crate that exists nowhere #60 — Publish the tdbin crate. docs/tdbin.html says tdbin = { path = "../typeDiagram/crates/tdbin" }; the crate is not on crates.io, so generated Rust does not compile outside the typeDiagram monorepo.
• [BLOCKER] tdbin rejects struct-form union variants — verify fails on the first union in a real schema #61 — Struct-form union variants are rejected. typediagram verify docs/models/live-ipc.td → error tdbin: variant 'OpenRange' must be bare or a single tuple field in v0. Aborts at the first union; four of our unions use this shape. Coverage of our schema is currently 0%.

Multi-language parity

• TypeScript tdbin codec: no CLI command, no List<T>, no scalar Option<T>, number not bigint #62 — TypeScript codec. No CLI command (Rust only), no List<T> — and every payload we send is list-shaped — no scalar Option<T>, and number instead of bigint for Int.
• Add Kotlin/JVM target (--to kotlin) + Kotlin tdbin runtime — the JVM is the only major ecosystem missing #63 — Kotlin/JVM target. --to has no Kotlin; the JetBrains plugin hand-rolls kotlinx.serialization against a schema it cannot generate from. --to kotlin has standalone value independent of tdbin.

Toolchain integration

• tdbin encode/decode ignore generation config — codecs don't compile against configured field types #64 — encode/decode ignore generation config. encode re-emits ADTs with no serde derives or configured field types; decode generates codecs that call .as_deref() on fields we override to Option<serde_json::Value>. Fixed properly by a native opaque-payload type (Add 'Any' / 'Json' field type for opaque payloads #28).
• --config cannot emit tdbin outputs — codecs go stale under --watch #65 — --config has no tdbin output. Outputs are restricted to converters.LANGUAGES, so codecs must be driven by separate shell-outs and go stale under --watch.

Order that unblocks us fastest

1. [BLOCKER] Publish the tdbin Rust runtime crate — generated codecs reference a crate that exists nowhere #60 + [BLOCKER] tdbin rejects struct-form union variants — verify fails on the first union in a real schema #61 — until both land, nothing else is testable end to end.
2. tdbin encode/decode ignore generation config — codecs don't compile against configured field types #64 (or Add 'Any' / 'Json' field type for opaque payloads #28, which subsumes most of it) — makes generated Rust codecs compile against our real types.
3. TypeScript tdbin codec: no CLI command, no List<T>, no scalar Option<T>, number not bigint #62 — second end of the wire; without it we maintain two encodings forever, which is worse than staying on JSON.
4. --config cannot emit tdbin outputs — codecs go stale under --watch #65 — makes it maintainable.
5. Add Kotlin/JVM target (--to kotlin) + Kotlin tdbin runtime — the JVM is the only major ecosystem missing #63 — brings the JVM client onto the same schema.

Related existing issues: #28 (opaque Json/Any type), #39 (field metadata), #41 (provenance header + --check), #51 (emitted-ADT template), #30 (pinned enum discriminants), #40 (payload-carrying unions lossy in C#).