---
id: LSN-0064
ticket: multi-frontend-sdk-canonical-definition
title: Runtime owns console ABI; PBS is delivery
created: 2026-09-19
tags: [compiler, compiler-general, sdk, stdlib, hostcalls, intrinsics, multi-frontend]
---

# Runtime owns console ABI; PBS is delivery

## Original Problem

"SDK" mixed three layers: platform ABI, PBS stdlib shells, and docs. The PBS
compiler reads `[Host]` and `[IntrinsicCall]` from `.pbs` files. It does not
read the runtime registry. Intrinsics already had a CSV in the runtime
repository, a handwritten Java enum, and a disabled parity test. Spec 8
documented `audio` and `@core:math`. Disk had `game/stdlib/1` with `color`,
`asset`, `composer`, `gfx`, `input`, and `log`. Spec 18 preferred `@sdk:*` for
host APIs while `@sdk:input` was a builtin.

There is no second language in this phase. The risk was not a missing generic
SDK. The risk was treating PBS shells, or a stale spec, as console ABI.
DEC-0047 had already classified canonical host identity as platform-owned and
left `PbsHostAdmissionValidator` in the PBS package until a shared source
existed.

## Consolidated Decision

This discussion was an audit. It published an authority map. It did not create
a canonical schema, generate bindings, enable intrinsic parity, or move the
host-admission validator.

The console has two distinct axes:

- VM-owned intrinsics, identity `canonical_name@version` plus `final_id`;
- host-backed calls, identity `(module, name, version)` plus capability.

Do not unify those axes into one catalog while PBS is the only frontend.

Conflict resolution:

- Intrinsic conflict: `runtime/docs/vm-arch/INTRINSICS.csv` wins. Studio
  `docs/vm-arch/INTRINSICS.csv` is a mirror. `IRVMIntrinsicDefinition` is
  derived. PBS `[BuiltinType]` / `[IntrinsicCall]` shells are delivery.
- Host conflict: runtime spec 16 and `SyscallMeta` win. PBS `[Host]` attributes
  must not define ABI. They declare the shell the PBS compiler reads today.

`@sdk:*` / `@core:*` paths must not imply ownership. `@sdk:input` is a
registered builtin exception. It was not moved to `@core`.

PBS conformance means PBS shells and docs match runtime ABI, not the reverse.
`PbsHostAdmissionValidator` may stay in the PBS package. `E_HOST_*` is not
eternal PBS semantics.

## Final Implementation

The work was editorial and test-facing.

Spec 18 now states that PBS stdlib is delivery and must not define syscall IDs,
intrinsic `final_id` values, or `SyscallMeta`. Runtime catalogs outrank it.
`@sdk:input` is recorded as the current builtin exception.

Spec 8 now documents `<target>/stdlib/<N>/...` with current line `1` at
`game/stdlib/1`: `core/color` and `sdk/{asset,composer,gfx,input,log}`. It no
longer ships `audio` or `@core:math` as current modules.

Spec 6.2 now requires PBS host shells to repeat the runtime registry and forbids
inventing `(module, name, version)`. The authority example is
`("gfx2d", "clear", 1)`.

`ConsoleAbiAuthoritySpecTest` locks those sentences. No production compiler
code, stdlib `.pbs`, or validator package moved. The intrinsic parity test
stays disabled.

## Examples

Runtime wins:

- `color.rgb@1` / `final_id` `0x1101` come from `INTRINSICS.csv`;
- `("gfx2d", "clear", 1)` and `("asset", "load", 1)` come from `SyscallMeta`.

PBS delivers, and does not own, those identities:

- `@core:color` with `[BuiltinType(name = "color", version = 1)]`;
- `@sdk:gfx` `LowGfx.clear` with `[Host(module = "gfx2d", name = "clear", version = 1)]`;
- `@sdk:input` with `[BuiltinType]` / `[IntrinsicCall]` under `@sdk`.

Stale docs that the recut removed as current line-1 inventory:

- `stdlib/1/sdk/audio`;
- `@core:math` as a packaged module.

## Pitfalls

- Do not treat PBS stdlib as the ABI catalog because the compiler already reads
  it. That is convenience, not authority.
- Do not create `HOSTCALLS.csv` or generate the Java enum because the map now
  exists. A catalog waits on a second consumer.
- Do not move `PbsHostAdmissionValidator` on the strength of this map. DEC-0047
  classified identity as platform; this discussion published who wins; the
  validator still has no shared catalog to consume.
- Do not read `@sdk` as host and `@core` as VM. Path is packaging. `@sdk:input`
  is builtin.
- Do not enable `IRVMIntrinsicRegistryParityTest` as a side effect of a doc
  recut. That is compiler/runtime sync, not PBS packaging conformance.
- Do not put the authority map only in a decision. Housekeep deletes workflow
  artifacts. Specs and this lesson keep the map.

## References

- Decision: `DEC-0048`
- Plan: `PLN-0127`
- Related lesson: `LSN-0063` - Language vs platform validation ownership
- Related lesson: `LSN-0061` - PBX is the language-neutral runtime boundary
- `docs/specs/compiler/18. Standard Library Surface Specification.md`
- `docs/specs/compiler-languages/pbs/8. Stdlib Environment Packaging and Loading Specification.md`
- `docs/specs/compiler-languages/pbs/6.2. Host ABI Binding and Loader Resolution Specification.md`
- `docs/vm-arch/README.md`
- `../runtime/docs/vm-arch/INTRINSICS.csv`
- `../runtime/docs/specs/runtime/16-host-abi-and-syscalls.md`
- `prometeu-compiler/prometeu-build-pipeline/src/test/java/p/studio/compiler/specs/ConsoleAbiAuthoritySpecTest.java`