AILang/docs/specs/0055-intrinsic-bodies.md

# Intrinsic bodies — Design Spec

**Date:** 2026-05-29
**Status:** Draft — awaiting user spec review
**Authors:** orchestrator + Claude

## Goal

Give Form-A a way to say *"this definition's body is supplied by
the compiler, not written in source"* — and make that the single,
honest representation for every definition whose real implementation
lives in the codegen intercept registry (`crates/ailang-codegen/src/intercepts.rs`,
shipped in raw-buf.1).

Today every such definition carries a **dummy body that never runs**.
The prelude's `eq Int` instance method is

```ail
(module dummy_today
  (fn eq_int_like
    (doc "Real implementation is a codegen intercept; the body false never executes.")
    (type (fn-type (params (con Int) (con Int)) (ret (con Bool))))
    (params x y)
    (body false)))
```

The `(body false)` (and the eighteen siblings like it across the
prelude) parses and type-checks, but codegen discards it and emits
the intercept instead. The body is a structural lie: a reader who
trusts the source is wrong about what runs. This is a standing
infraction of the honesty rule (`design/contracts/0007-honesty-rule.md`),
which the language has carried silently since the operator-routing
milestone.

The lie is not merely cosmetic. It blocks the `raw-buf` milestone:
a polymorphic kernel-tier function such as `RawBuf`'s `get : RawBuf a -> Int -> a`
needs a placeholder body that produces a value of type `a` — and
AILang has **no value of polymorphic type**. There is no honest
expression to write. The dummy-body requirement makes the function
structurally impossible to author, which is what BLOCKED raw-buf.2
(see `git log` 4ad003d..647121c and the discarded .2 working tree).

The fix is a new Form-A surface affordance, `(intrinsic)`, that
replaces the `(body ...)` clause for compiler-supplied definitions.
This is the same construct every systems language has: LLVM `declare`,
Rust `extern "rust-intrinsic"` / `#[rustc_intrinsic]`, Haskell
`foreign import prim`, C compiler builtins. The intercept registry
*is* AILang's compiler-supplied-body table; `(intrinsic)` is the
surface declaration of membership in it.

## Architecture

Three landing points, two iterations.

**Iteration `intrinsic-bodies.1` — the mechanism.**

1. **AST.** A new leaf term variant `Term::Intrinsic` is the body of
   a compiler-supplied definition. `FnDef.body` and `Term::Lam.body`
   keep their existing types (`Term` / `Box<Term>`) — they are not made
   optional. A definition is intrinsic iff its body *is*
   `Term::Intrinsic`. The variant is strictly additive: a fixture that
   does not use it serialises bit-identically (the established
   additive-term-variant pattern — `Term::New`, `Term::Loop`,
   `Term::Recur`, `Term::Clone`, `Term::ReuseAs` all landed this way,
   `design/contracts/0002-data-model.md`). This is the load-bearing
   representation choice: it puts "this body is compiler-supplied" at
   the body position where it semantically belongs, keeps the
   ~150-site `body`-read/construct surface across six crates untouched
   (only exhaustive `match`-on-`Term` arms gain a case), and makes the
   illegal "body AND intrinsic" state *unrepresentable* rather than
   rejected — a body is either `Term::Intrinsic` or a real term, never
   both. (The rejected alternative, `body: Option<Term>` + an
   `intrinsic: bool` flag, splits one fact across two fields, admits
   that illegal state, and breaks every `body` consumer in the
   workspace.)
2. **Form-A surface.** For a top-level fn, the parser accepts
   `(intrinsic)` as a sibling clause where `(body ...)` would go and
   maps it to `body = Term::Intrinsic`; the printer emits `(intrinsic)`
   in that slot when the body is `Term::Intrinsic`. For a lambda,
   `(intrinsic)` sits at the positional body slot (where the body term
   goes today) and parses to `Term::Intrinsic`; the printer emits it
   there. Round-trip (`parse ∘ print = id`) holds by the standard
   fixture gate. The fn-vs-lam surface asymmetry mirrors the existing
   one (a fn carries a named `(body X)` clause; a lambda carries its
   body positionally) — `(intrinsic)` follows each form's existing body
   placement.
3. **Checker.** A definition whose body is `Term::Intrinsic`
   type-checks against its *signature only* — there is no body term to
   infer. One new reject: an intrinsic definition outside a
   `(kernel)`-tier module or the prelude is rejected with
   `intrinsic-outside-kernel-tier`. (There is no `intrinsic-with-body`
   reject — the representation makes that state impossible, see point 1.)
4. **Codegen.** A definition whose body is `Term::Intrinsic` routes
   through `intercepts::lookup`; if no intercept is registered for its
   mangled name, codegen emits the existing deferral diagnostic. The
   current "lower the body" path is not taken for these definitions —
   `lower_term` is never called on a `Term::Intrinsic` body (and a
   `Term::Intrinsic` reaching `lower_term` through any other path is a
   codegen-internal error, since an intrinsic body must be consumed by
   the intercept route).
5. **Ratifier.** One throwaway intrinsic in the `kernel_stub`
   fixture — a nullary `answer : () -> Int` whose intercept emits
   `ret i64 42` — drives the mechanism end-to-end (parse → check →
   codegen → run). `kernel_stub` already exists precisely to ratify
   kernel-tier mechanisms.

**Iteration `intrinsic-bodies.2` — the migration + the lock.**

6. **Prelude migration.** The **thirteen authored dummy bodies** in
   `examples/prelude.ail` swap their inner `(body <dummy>)` for
   `(intrinsic)`: the seven Eq/Ord instance methods (`eq` for
   Int/Bool/Str/Unit, `compare` for Int/Bool/Str) and the six `float_*`
   free fns (`float_eq/ne/lt/le/gt/ge`). These are exactly the
   `INTERCEPTS` entries that today carry a placeholder body a reader
   would mistake for the implementation.

   **Not migrated, and why** (the count reconciliation — the registry
   has nineteen entries, not eighteen-all-dummies):
   - `lt__Int`, `le__Int`, `gt__Int`, `ge__Int`, `ne__Int` (5) intercept
     the *monomorphised* `__Int` specialisations of the polymorphic free
     fns `lt/le/gt/ge/ne`, which carry **real bodies** in the prelude
     (`(match (app compare x y) …)`, `(app not (app eq x y))`). Those
     bodies are honest and live — they are lowered for every non-`Int`
     instantiation; only the `Int` specialisation is intercepted for a
     faster direct `icmp`. These are an **optimisation class**, not a
     compiler-supplied-body class: there is no lie to fix and no source
     body to replace, so they get **no** `(intrinsic)` marker.
   - `answer` (1) is the `.1` ratifier, already `(intrinsic)` in
     `STUB_AIL` — not a prelude site.

   So: 13 prelude migrations + 5 optimisation-class entries + 1 already
   migrated = 19 `INTERCEPTS` entries.
7. **Hard-lockstep pin.** The registry splits into two classes: an
   entry is **intrinsic-backed** when a `(intrinsic)` marker in the
   loaded workspace resolves to its mangled name (the 13 prelude
   migrations + `answer` = 14), or **optimisation-only** when it
   intercepts a real-bodied fn's specialisation (the 5 `*__Int` icmp
   family). The pin asserts a **bijection over the intrinsic-backed
   class**: every workspace `(intrinsic)` marker has exactly one
   `INTERCEPTS` entry, and every `INTERCEPTS` entry *not* on the
   explicit optimisation-only allowlist has exactly one workspace
   marker. The optimisation-only set is an explicit, documented
   allowlist in the pin (the 5 names); adding a registry entry that is
   neither marker-backed nor allowlisted is a red test, as is an
   intrinsic marker with no entry. This extends the
   `registry_contains_all_legacy_arms` pin from raw-buf.1 from a
   one-directional name check to a two-directional source↔registry
   bijection.
8. **Dead-path confirmation.** `.1` already routes every intercepted
   def by name (`try_emit_primitive_instance_body` consults
   `intercepts::lookup` on the fn name *before* the body is inspected),
   so a prelude dummy body is bypassed before it could reach
   `lower_term` — the dummy is already dead, not lowered-then-discarded.
   After migration the dummy is gone entirely (replaced by
   `Term::Intrinsic`). This point is therefore a **verification, not a
   removal**: confirm no codegen path lowers an intercepted body, and
   delete now-stale comments that reference the discarded dummy. The
   `Term::Intrinsic` internal-error guards in `lower_term` / the synth
   walker (added in `.1`) STAY — they guard against an intrinsic body
   escaping its definition slot, which is a different concern.

## Concrete code shapes

### North-star: what a kernel author writes (proposed surface)

The surface this milestone delivers. These blocks are labelled
`scheme` (not `ail`) on purpose — the `(intrinsic)` attribute does
**not** parse against the pre-feature tool, so they are proposals,
not ratified current behaviour. (Verified: see the must-fail
transcript below.)

The `raw-buf` north-star — a polymorphic kernel-tier function with
no honest body, the case that motivated the milestone:

```scheme
(fn get
  (doc "Indexed read. Codegen intercept get__RawBuf__<T> emits a getelementptr + load.")
  (type (forall (vars a) (fn-type (params (borrow (con RawBuf a)) (con Int)) (ret (con a)))))
  (params b i)
  (intrinsic))
```

The prelude `eq Int` instance method after migration (the lie, fixed).
The marker sits on the *lambda body*; the lambda's typed shell
(`params`/`ret`) stays — it is the method's local signature, and
intrinsic drops the body, not the signature (see the rationale under
"Implementation shape" below):

```scheme
(instance
  (class Eq)
  (type (con Int))
  (doc "Eq Int. Codegen intercept eq__Int emits icmp eq i64 with alwaysinline.")
  (method eq
    (body (lam (params (typed x a) (typed y a)) (ret (con Bool)) (intrinsic)))))
```

The `intrinsic-bodies.1` ratifier — the throwaway smoke intrinsic in
the `kernel_stub` fixture:

```scheme
(fn answer
  (doc "Ratifies the intrinsic mechanism end-to-end. Intercept answer emits ret i64 42.")
  (type (fn-type (params) (ret (con Int))))
  (params)
  (intrinsic))
```

### Current-state evidence (parses today)

The dummy-body lie exactly as the prelude carries it now — a body
that parses, type-checks, and never runs:

```ail
(module dummy_today
  (fn forty_two
    (doc "Real implementation is a codegen intercept. The body 0 never executes.")
    (type (fn-type (params (con Int)) (ret (con Int))))
    (params n)
    (body 0)))
```

### Must-fail evidence (rejected today)

`(intrinsic)` is not yet a surface attribute, so the proposed forms
above are rejected by the pre-feature tool. This is the baseline the
milestone moves off — the parser reject is what `intrinsic-bodies.1`
turns into an accept (inside kernel-tier / prelude) and a *different*
reject (`intrinsic-outside-kernel-tier`, for user modules):

```console
$ cat intr.ail
(module intr
  (fn new (type (fn-type (params (con Int)) (ret (con Int)))) (params n) (intrinsic)))
$ ail check intr.ail
error: [surface-parse-error] parse error in fn-def: unknown fn attribute
  `intrinsic`; expected `doc`, `export`, `suppress`, `type`, `params`,
  or `body` at byte 86
```

### Implementation shape (secondary — the AST delta)

The change is one new leaf `Term` variant. Exact bytes are the
planner's job; this fixes the shape only.

`Term::Intrinsic` (`crates/ailang-core/src/ast.rs`):

```jsonc
// new leaf term — the body of a compiler-supplied definition.
// Strictly additive (no skip_serializing_if needed; a fixture that
// does not carry it hashes bit-identically — the same additive-variant
// pattern as Term::New / Term::Loop / Term::Recur / Term::Clone).
// Never reduces to a value: codegen consumes it via intercepts::lookup,
// the typechecker treats a def with this body as signature-only.
{ "t": "intrinsic" }
```

`FnDef.body` (currently `Term`) and `Term::Lam.body` (currently
`Box<Term>`) are **unchanged in type** — they simply may now hold
`Term::Intrinsic`. A top-level intrinsic fn is
`FnDef { body: Term::Intrinsic, .. }`; an intrinsic instance-method
lambda is `Term::Lam { body: Box::new(Term::Intrinsic), .. }`. "Is
this definition intrinsic?" is `matches!(body, Term::Intrinsic)`.

The marker sits at the body position — for a top-level fn that is the
fn body; for an instance method that is the *lambda body*, **not** the
method as a whole — and this placement is load-bearing, not
incidental.

The reason is local reasoning (`design/INDEX.md` § Goal: "every
definition carries its full type and effect set, so a signature can
be trusted without reading the body"). An intrinsic is a *signature
without a body*, never *nothing* — exactly as LLVM `declare i64
@llvm.foo(i64, i64)`, Rust `extern "rust-intrinsic" { fn foo(x: i32)
-> i32; }`, and Haskell `foreign import prim` all keep the full
signature and drop only the body. For a top-level fn the signature is
the `(type ...)` clause, which stays beside the `(intrinsic)` marker.
For an instance method the signature-bearing structure is the lambda
itself: its `(params (typed x a) (typed y a))` and `(ret (con Bool))`
are the parameter types and return type, readable at the definition
site. Hoisting the marker to the method (`(method eq (intrinsic))`)
would erase that local signature — a reader would have to climb to
the `Eq` class declaration and substitute `a := Int` mentally to
recover it. So the marker is the lambda's *body* (`Term::Intrinsic`),
leaving the lambda's typed shell — the local signature — in place. The
mono pass (`crates/ailang-check/src/mono.rs::synthesise_mono_fn`)
already reads the method's parameter names and inner body out of this
lambda (`Term::Lam { params, body, .. } => (params, *body)`); with the
inner body being `Term::Intrinsic`, the synthesised `FnDef` carries
`body: Term::Intrinsic` and is itself intrinsic. The destructure
itself is unchanged — only what the inner body *is* differs.

Hash-stability note: `Term::Intrinsic` is a new enum variant, not a
new field, so a fixture that does not use it serialises exactly as
before — the same mechanism that kept hashes stable when `Term::New`,
`Term::Loop`, and `Term::Recur` were added. The `design_schema_drift.rs`
schema mirror, the `schema_coverage.rs` variant corpus (a fixture must
now exercise `Term::Intrinsic`), and the `0002-data-model.md` contract
all move in the same iteration as the variant.

## Components

| Component | Iteration | Change |
|---|---|---|
| `crates/ailang-core/src/ast.rs` | .1 | New leaf variant `Term::Intrinsic`. `FnDef.body` / `Lam.body` types unchanged. |
| `crates/ailang-core` canonical/hash/visit | .1 | Exhaustive `match`-on-`Term` arms gain a `Term::Intrinsic` case (leaf, no sub-terms to walk). Schema-coverage corpus extended to exercise it. |
| `crates/ailang-surface` (lex/parse/print) | .1 | `(intrinsic)` parsed + printed: as a body-slot clause in `fn-def`, at the positional body slot in `lam`; maps to/from `Term::Intrinsic`; round-trip gated. |
| `crates/ailang-check/src/lib.rs` | .1 | A def whose body is `Term::Intrinsic` checks signature-only; rejects intrinsic outside kernel-tier/prelude (`intrinsic-outside-kernel-tier`). No `intrinsic-with-body` reject — the representation forbids that state. |
| `crates/ailang-check/src/mono.rs` | .1 | `synthesise_mono_fn` lambda-destructure unchanged; an inner `Term::Intrinsic` body flows through to an intrinsic synthesised `FnDef`. |
| `crates/ailang-codegen/src/lib.rs` | .1 | A def whose body is `Term::Intrinsic` routes through `intercepts::lookup`; `lower_term` is never called on it. |
| `crates/ailang-kernel-stub` + `ailang-surface` parse hop | .1 | `answer` smoke intrinsic added to the stub fixture; its intercept registered. |
| `examples/prelude.ail` | .2 | 13 authored dummy bodies → `(intrinsic)` (7 Eq/Ord instance methods + 6 `float_*` fns). The 5 `*__Int` icmp-family and `answer` are NOT prelude sites. |
| `crates/ailang-codegen/src/intercepts.rs` (pin) | .2 | `registry_contains_all_legacy_arms` upgraded to a bijection over the intrinsic-backed class + an explicit 5-name optimisation-only allowlist. Pin loads the workspace + monomorphises to recover mangled names for the marker direction. |
| `crates/ailang-surface/tests/prelude_module_hash_pin.rs` | .2 | Rebaseline the prelude module hash (13 bodies change). |
| `crates/ail/tests/mono_hash_stability.rs` | .2 | Rebaseline the 6 post-mono `eq__*`/`compare__*` def-hashes (body flips to `Term::Intrinsic`). The 4 `show__*` pins do NOT move. |
| codegen dummy-body path | .2 | Already dead post-`.1` (intercept-by-name precedes body inspection); `.2` confirms + deletes stale comments, no behavioural removal. |
| `design/contracts/0002-data-model.md` | .1 | New `{ "t": "intrinsic" }` Term entry; note in `fn`/`lam` that the body may be `Term::Intrinsic`. |
| `design/contracts/0007-honesty-rule.md` | .2 | Edit only if it names the prelude dummies; recon found it is a general rule that may not reference them — flag honestly, no forced edit. |

## Data flow

Authoring → parse → AST → check → codegen, unchanged in topology;
`Term::Intrinsic` is recognised at three of those stations:

1. **Parse.** `(intrinsic)` in a fn body slot or a lambda body slot
   produces a `Term::Intrinsic` body. There is no "both body and
   intrinsic" surface form to reject — the grammar offers one body
   slot, and `(intrinsic)` either fills it or it does not.
2. **Check.** A def whose body is `Term::Intrinsic`: validate the
   signature, skip body inference. Enforce the kernel-tier/prelude
   scope. The module's `kernel: true` flag (or prelude identity) is
   already available to the checker via the loaded workspace.
3. **Codegen.** A def whose body is `Term::Intrinsic`:
   `intercepts::lookup(mangled_name)`. Hit → emit the intercept.
   Miss → the existing deferral diagnostic (same one raw-buf.2 reuses
   for unregistered `RawBuf` ops).

## Error handling

| Condition | Diagnostic | Station |
|---|---|---|
| `(intrinsic)` in a non-kernel, non-prelude module | `intrinsic-outside-kernel-tier` | check |
| Intrinsic def with no registered intercept | existing codegen deferral | codegen |

The first row is the honesty-rule guard at the workspace boundary:
user code cannot mark a body as compiler-supplied, so the lie cannot
re-enter through user modules. (There is deliberately no
"body-and-intrinsic" row — `Term::Intrinsic` is a body, so a def
either has it or has a real body, never both. The illegal state is
unrepresentable, not diagnosed.)

## Testing strategy

**intrinsic-bodies.1:**

- Round-trip: a kernel-tier fixture carrying `(intrinsic)` on both a
  top-level fn and a lambda parses → prints → re-parses to canonical-byte
  equality (rides the existing `round_trip.rs` corpus gate).
- Check accept: the `kernel_stub` `answer` intrinsic checks clean.
- Check reject (scope): a user module with an `(intrinsic)` fn is
  rejected with `intrinsic-outside-kernel-tier`.
- E2E: `answer` builds and runs, exit/print observing `42` — the
  mechanism works from source to native.
- Schema drift: `design_schema_drift.rs` green against the new
  `0002-data-model.md`; `schema_coverage.rs` observes `Term::Intrinsic`
  in the fixture corpus.

**intrinsic-bodies.2:**

- Hard-lockstep pin: bijection over the intrinsic-backed class. Add an
  intrinsic marker with no `INTERCEPTS` entry → red; add a
  non-allowlisted `INTERCEPTS` entry with no marker → red; drop either
  → red. The 5-name optimisation-only allowlist is explicit in the pin;
  the pin loads the workspace, monomorphises, and walks post-mono
  `FnDef` bodies for `Term::Intrinsic` to recover mangled names for the
  marker direction.
- All pre-existing E2E ratifiers stay green (the eq/compare/float
  smoke suite named in raw-buf.1's commit body:
  `eq_primitives_smoke_compiles_and_runs`,
  `compare_primitives_smoke_prints_1_2_3_thrice`,
  `float_compare_smoke_prints_true_true_false`,
  `eq_ord_polymorphic_runs_end_to_end`) — the migration is
  behaviour-preserving (the intercepts emit the same IR; only the
  now-dead source body changes).
- Hash rebaselines (both move; both recorded old→new in the commit
  body): `prelude_module_hash_pin.rs` (prelude module hash; 13 bodies
  change) and `mono_hash_stability.rs` (the 6 post-mono
  `eq__*`/`compare__*` def-hashes flip to `Term::Intrinsic`; the 4
  `show__*` pins do NOT move). IR snapshots do NOT change (no prelude
  instance method is codegen'd into the user-program snapshots).

## Acceptance criteria

Applied prospectively (`design/contracts/0004-feature-acceptance.md`):

- **An LLM author naturally reaches for it.** A kernel author writing
  `RawBuf.get` has no honest body to write — there is no value of type
  `a`. Faced with the dummy-body requirement, the natural move is to
  declare "the compiler supplies this", which is exactly `(intrinsic)`.
  The north-star block above *is* that program; it is the empirical
  evidence, not a prose assertion.
- **Measurably improves correctness / removes redundancy.** Closes a
  standing honesty-rule infraction (eighteen bodies that lie about
  what runs); removes the dead body-lowering path for intercepted
  defs; removes the structural impossibility blocking polymorphic
  kernel-tier functions.
- **Reintroduces no eliminated failure class.** Machine-readability,
  local reasoning, provability, hallucination-robustness all hold or
  improve: a reader of an `(intrinsic)` def now knows from the source
  alone that the body is compiler-supplied, instead of having to read
  the codegen table to discover the source was lying.

Out of scope: a user-facing plugin API for registering custom
intercepts (the scope guard explicitly forbids `(intrinsic)` in user
modules); any change to the intercept *dispatch* mechanism shipped in
raw-buf.1; the `raw-buf` redo itself (a separate milestone that
consumes this one).