Wire envelope: split the single u8 discriminant into (trait, method) u8 pair

[decrypto21]

Context

Agreed in the SDK team chat (PG + Sergey, 2026-08-06): change the TrUAPI wire envelope's method identifier from a single u8 to two u8s — one for the namespace (the API trait), one for the method within that trait.

This is a wire-breaking change. It will not merge until every host runs the Rust-based implementation; until then the branch is kept rebased and ready.

Current state

The envelope is [SCALE str requestId][u8 discriminant][payload] (js/packages/truapi/src/transport.ts, truapi-server/src/frame.rs on the PR-104/PR-295 lineage). The single u8 is the whole address:

• Ids are hand-written #[wire(request_id = N)] attributes in rust/crates/truapi/src/api/*.rs; a request burns 2 slots, a subscription 4. They are append-only and never reused.
• 163 of 255 values are already assigned; ids 70–75 are permanently burned by the retired JsonRpc trait, and coin_payment alone consumed 28 slots.
• Five traits (notifications, account, signing, statement_store, payment) already have non-contiguous id blocks because new methods can only append at the global tail.

Splitting into (trait: u8, method: u8) gives each trait its own 256-slot method space, keeps trait id blocks contiguous forever, and makes the namespace visible on the wire for dispatch and debug tooling.

Proposed change

New envelope: [SCALE str requestId][u8 trait][u8 method][payload].

• Trait ids are assigned explicitly (new trait-level wire attribute), pinned forever.
• Method ids restart from 0 within each trait. Since this ships as a coordinated cutover, existing method ids are renumbered once; append-only then resumes per trait.
• Handshake codec version bumps 1 → 2. The handshake itself rides the changed envelope, so old↔new peers cannot negotiate — the point of the bump plus loud unknown-discriminant handling is that a skewed pair fails with a visible error instead of the silent drop + infinite "Handshaking" hang we shipped in the 0.6↔0.7 window (Wire regression between @novasamatech/host-api@0.8.3 and 0.8.4: silent handshake drop, misleading 'Host doesn't support it' log triangle-js-sdks#200 is the same failure mode one protocol layer up).

Scope

In truapi:

• truapi-macros: trait-level namespace id + per-method ids in #[wire(...)]; keep the rustdoc doc-smuggling path working
• truapi-codegen: extraction (rustdoc.rs), sort/collision/overflow logic and TS emitters (ts.rs), Rust emitters (rust/wire_table.rs, dispatcher) — collision detection becomes per-trait
• Envelope codecs: js/packages/truapi/src/transport.ts encode/decode, truapi-server/src/frame.rs
• Dispatch: truapi-server/src/dispatcher.rs and truapi-host TS dispatch keyed on the pair
• Id-keyed debug surfaces: SENSITIVE_FRAME_IDS, WIRE_DECODE_TABLE, truapi-debugger frame decoding
• Re-pin byte-exact tests: wire-equality.test.mjs golden hex, wire-table-loop.test.mjs, frame.rs frame-length asserts, codegen golden files
• Codec version 2 + explicit error on unknown (trait, method) on both sides

Follow-up, separate PR, only if it executes quickly (per PG): consume @parity/truapi's generated wire table inside triangle-js-sdks instead of the hand-maintained allocator in packages/host-api/src/protocol/impl.ts — that table has already diverged from the spec (max id 135 vs 163), and triangle-js-sdks is retired once hosts move to the Rust SDK.

Sequencing

• Base after feat: implement RFC-0024 ring VRF key management #348 lands (it adds new wire ids on feature/people-rfc); the renumbering picks them up either way.
• Merge gate: all hosts on the Rust implementation.

Open questions

• Trait id assignment order — proposal: explicit literal per trait, starting from the current api/mod.rs order, pinned in the spec.
• Renumber method ids per trait from 0 (proposed) vs. preserving current global numbers in the method byte.
• Confirm codec-version bump to 2 as the cutover gate.