AILang/design/models/0007-kernel-extensions.md

# Kernel extensions — plugin-style domain types whitepaper

**STATUS.** Mechanisms milestone closed 2026-05-28
(`kernel-extension-mechanics`; iterations prep.1, prep.2, prep.3
landed). The four language-level mechanisms — type-scoped
namespacing, `Term::New`, kernel-tier modules + auto-import, and
`param-in` — are shipped and ratified by the stub kernel crate
(`crates/ailang-kernel-stub/`). The base-extension and library-
extension milestones (`raw-buf`, `series`) are pending. This
whitepaper describes the design as a coherent whole; the
per-milestone specs in `docs/specs/` carry the implementation-
level detail. Sections describing the now-shipped mechanisms read
in present-state per the
[honesty-rule](../contracts/0007-honesty-rule.md); sections about
the still-pending milestones keep their forward-looking framing.

## The problem

AILang's core needs to stay small. Five primitive types (`Int`,
`Bool`, `Unit`, `Str`, `Float`), algebraic data types, typeclasses,
algebraic effects, RC + uniqueness — that is the language. New
primitive types are a tax: they inflate the schema, the checker,
the codegen, the documentation surface, and the LLM-author's
mental model of what the language is.

But: AILang authors will work in specific domains. Streaming
analytics wants a bounded ring buffer. Numerical work wants flat
arrays and matrices. Financial work wants fixed-precision decimals
and timestamps with arithmetic. None of these belong in
`ailang-core` — yet without them, the LLM author has to hand-roll
linked-list approximations that are algorithmically wrong (O(N²)
where O(N) is the natural complexity).

The resolution: **kernel extensions** — a plugin contract that
lets a domain-specific type live in its own crate, register with
the compiler at boot time, and present itself to user code as if
it were a built-in primitive. Core has no domain knowledge of any
specific extension; extensions have no core dependency beyond the
plugin contract.

The first consumer is `Series` — a bounded ring buffer for
streaming data. Future consumers are not specified here but the
shape is intentional: each new domain type is one new crate, zero
core changes.

## Worked example — `Series` for streaming analytics

What an LLM-author writes when given the task "compute simple
moving average over a stream of float values, window size 3":

```
(module sma_demo

  (data FloatList
    (ctor FNil)
    (ctor FCons (con Float) (con FloatList)))

  (fn sum_window_step
    (type (fn-type
      (params (borrow (Series (con Float))) (con Int) (con Int) (con Float))
      (ret (con Float))))
    (params s n i acc)
    (body
      (if (app eq i n)
        acc
        (tail-app sum_window_step s n
          (app + i 1)
          (app + acc (app Series.at s i))))))

  (fn emit_if_full
    (type (fn-type
      (params (borrow (Series (con Float))) (con Int))
      (ret (con Unit)) (effects IO)))
    (params s n)
    (body
      (if (app >= (app Series.total_count s) n)
        (app print (app /
          (app sum_window_step s n 0 0.0)
          (app int_to_float n)))
        unit)))

  (fn run_stream
    (type (fn-type
      (params (own (Series (con Float))) (con Int) (con FloatList))
      (ret (con Unit)) (effects IO)))
    (params s n input)
    (body
      (match input
        (case (pat-ctor FNil) unit)
        (case (pat-ctor FCons v rest)
          (let s_new (app Series.push s v)
            (seq (app emit_if_full s_new n)
              (tail-app run_stream s_new n rest)))))))

  (fn main
    (type (fn-type (params) (ret (con Unit)) (effects IO)))
    (params)
    (body
      (let s (new Series (con Float) 3)
        (app run_stream s 3
          (term-ctor FloatList FCons 1.0
          (term-ctor FloatList FCons 5.0
          (term-ctor FloatList FCons 3.0
          (term-ctor FloatList FCons 8.0
          (term-ctor FloatList FCons 6.0
          (term-ctor FloatList FCons 2.0
          (term-ctor FloatList FNil)))))))))))
```

Four constructs in this program are not in today's AILang. Each
maps onto one of the four mechanisms this whitepaper defines:

- `Series` appears in type position with no module qualifier
  (`(con Series (con Float))`) — *kernel-tier modules*.
- `(new Series (con Float) 3)` constructs the series — *new term
  construct*.
- `(app Series.at s i)`, `(app Series.total_count s)`,
  `(app Series.push s v)` — *type-scoped namespacing*.
- The `(own (Series ...))` mode on `run_stream` plus the linear
  state-threading via `(let s_new ...)` is the *uniqueness mode*
  discipline that already exists in AILang. Series carries no
  separate algebraic effect; its mutation is mode-tracked, not
  effect-tracked. Under uniqueness inference (Gitea #22), the
  rebuild-on-push compiles to in-place mutation.

A fifth mechanism, `param-in`, is not visible at the call site but
enforces that `(con Series (con Str))` would be rejected.

## The four mechanisms

### 1. Type-scoped namespacing

**Form.** `<TypeName>.<member>` resolves to the `<member>`
definition in the home module of `<TypeName>`. The home module of
a type `T` is the module whose `defs` contains a `TypeDef` named
`T`.

**Canonical form.** Type-scoped access is the *canonical* form for
operations associated with a type. The pre-existing module-scoped
form (`<module>.<member>`) remains a valid surface only for
free-standing definitions not bound to a specific type receiver,
and as a collision disambiguator if two types in the workspace
ever shared a name (which is not the case today).

**Examples.**

```
(app Series.at s i)               ; series_at via type-scope
(app Maybe.from_maybe 0 m)        ; from_maybe via type-scope  (was: std_maybe.from_maybe)
(con Series (con Float))          ; type itself accessible bare under kernel-tier
(con Maybe (con Int))             ; type itself, if Maybe is in scope via import
```

**Why this is LLM-natural.** When asked "what can I do with a
Series?", an LLM produces `Series.<x>` queries. When asked "what
can I do with a Maybe?", it produces `Maybe.<x>` queries. The
type, not the module, is the LLM's mental anchor. Module-scoped
access requires knowing where a function lives — extra cognitive
load with no expressive benefit.

**Schema impact.** Zero. The dotted form already parses today as
`Term::Var { name: "Maybe.from_maybe" }` (the `.` is a permitted
identifier character in the existing lexer). What changes is the
*resolver*: it tries the type-namespace branch first, falls back
to the module-namespace branch.

**Diagnostics.** `TypeScopedMemberNotFound` (the receiver is a
known type, but the member is not defined in its home module);
`TypeScopedReceiverNotAType` (the receiver name is neither a
module nor a type).

### 2. The `new` term construct

**Form.** `(new T arg1 arg2 ...)` where `T` is a `TypeName` and
`argN` is either a `Type` expression (e.g. `(con Float)`) or a
`Term` expression. The construct desugars to: look up `new` in the
home module of `T`, call it with the given args.

**AST.** New variant `Term::New { type_name: TypeName, args: Vec<NewArg> }`
where `NewArg = NewArg::Type(Type) | NewArg::Value(Term)`. Each arg
carries its kind in serialised JSON.

**Why complementary to `term-ctor`, not replacing it.** `term-ctor`
constructs an ADT value via a *named data constructor* (a tag in
the TypeDef's `ctors` list). `new` calls a *function* in the
type's home module. They serve different purposes:

- `(term-ctor IntList ICons 5 INil)` builds an `IntList` from its
  `ICons` ctor. This is structural — the AST node carries the
  ctor name and field args directly.
- `(new Series (con Float) 3)` calls `Series.new : (Type, Int) -> Series a`.
  This is functional — there is no `Series` data ctor; the
  type is opaque and construction is a defined operation.

A type can have both: an ADT could expose `term-ctor` for direct
ctor access plus a `new` def for builder-style construction. The
two paths are not in tension; they answer different design needs.

**Diagnostics.** `NewTypeNotConstructible` (no `new` def in the
type's home module); `NewArgKindMismatch` (a Type arg where the
signature expects a Value, or vice versa).

### 3. Kernel-tier modules with auto-import

**Form.** A new boolean field `kernel` on `Module` (omitted when
false, so existing modules hash-stable). When `kernel: true`:

- The module's top-level defs are accessible bare in every other
  module of the workspace, with no `(import ...)` declaration
  required.
- The module's types are accessible bare in type position.

**Why.** The architectural property the user wants is: the domain
type lives outside core, but at the call site it feels like a
primitive. Bare `Series` in `(con Series (con Float))` without
seeing `(import series)` somewhere is the ergonomic property.

**Single mechanism.** Auto-injection is a single load-time hop,
generalised from a hardcoded single-name path (prelude only,
pre-prep.3) to a flag-driven multi-module mechanism (since
prep.3 of `kernel-extension-mechanics`). The loader at
`crates/ailang-surface/src/loader.rs::load_workspace` injects
`parse_prelude()` and `parse_kernel_stub()` programmatically,
then derives the implicit-imports list from
`modules.values().filter(|m| m.kernel)`. Every module in the
workspace's modules map that carries `kernel: true` enters the
auto-import set; consumers see those modules' top-level types
bare without an `(import …)` declaration.

Prelude carries `kernel: true` in `examples/prelude.ail`. Its 12
free fns (see `examples/prelude.ail:85-148`) remain callable bare
in every consumer module — ratified by
`crates/ail/tests/prelude_free_fns.rs` across the code-path
migration. The ratifying stub kernel module
(`crates/ailang-kernel-stub/`) and any future kernel-tier
consumers (Series, Matrix) auto-import through the same filter.

Class-method dispatch (a separate mechanism — see
[method dispatch](../contracts/0016-method-dispatch.md)) is
orthogonal: the dispatch is about how method calls are
resolved; the auto-import is about which names are in scope.

### 4. `param-in` — closed-set type parameter restriction

**Form.** Optional field on `TypeDef`: `param-in: Map<VarName, Set<TypeName>>`.
Skipped when absent. Each entry constrains a type variable to a
fixed set of named types.

**Example.**

```jsonc
// TypeDef Series
{ "kind": "type",
  "name": "Series",
  "vars": ["a"],
  "ctors": [],
  "param-in": { "a": ["Int", "Float", "Bool"] }
}
```

When the checker sees `Type::Con { name: "Series", args: [Type::Con { name: "Str", ... }] }`,
it looks up `Series`'s `param-in`, finds that `a` must be in
`{Int, Float, Bool}`, and rejects `Str`.

**Why not a marker class.** `class Primitive a` with three
instances (Int, Float, Bool) and no methods would express the
same restriction via existing typeclass machinery. Reason it is
not chosen: class constraints model *behaviour* (a type satisfies
some interface). `param-in` models *structural identity* (the
type's storage layout). Series's element-type restriction is
structural — the C-runtime ring buffer has a `data[]` whose layout
depends on whether the element is 8 bytes (Int/Float) or 1 byte
(Bool). A marker class would model this as "Int is a Primitive"
which is true but indirect. `param-in` says directly: "Series's
element variable must be one of these three named types". The
typecheck site is a set-membership lookup, not a class-instance
discharge.

**Diagnostic.** `ParamNotInRestrictedSet` names the offending
type and the allowed set.

## Two-tier extension architecture

Kernel extensions split into two tiers, distinguished by what
they need to provide:

### Base extensions

Provide primitives that are *not expressible in AILang itself* —
typically because they require mutable indexed storage,
hardware/OS interaction, or library-wrapping that has no AILang
surface. A base extension ships:

1. **A kernel-tier module manifest** with the TypeDef
   (and `param-in` if applicable) plus the operation signatures.
2. **Codegen intercepts** — Rust code registered with the
   compiler that, at each call site of an operation, emits the
   appropriate LLVM IR text (using existing runtime symbols like
   `@ailang_rc_alloc` for allocation). This follows the
   `try_emit_primitive_instance_body` precedent: Rust returns IR
   text, no separate C glue.
3. **Optional C runtime support** in `runtime/<extension>.c` —
   only when the operations *truly cannot be expressed as IR
   over existing runtime primitives* (e.g. wrapping PCRE2 for a
   future Regex extension). For the foreseeable base extensions
   (RawBuf, Matrix), no C is needed.

The first base extension is `RawBuf T` — a mutable, indexed,
bounded-size flat buffer of primitive elements. Element type is
restricted to `{Int, Float, Bool}` via `param-in` for the MVP.

### Library extensions

Provide domain-specific types that *are* expressible in AILang
once a base extension has been added. A library extension ships:

1. **A kernel-tier `.ail` module** (or `.ail.json`) — plain
   AILang code, no Rust intercepts, no C. The module declares an
   ADT and a set of fns; the fns call into a base extension's
   API.
2. **Nothing else.** No codegen intercepts (regular AILang
   codegen handles it), no C runtime (regular AILang RC handles
   it), no compiler-side plug-ins.

The first library extension is `Series T` — a bounded ring
buffer with financial-style indexing. Implemented as an ADT
wrapping a `RawBuf T` plus four bookkeeping `Int` fields
(lookback, head, count, total). All operations are AILang fns.

### Why split

- **AILang-in-AILang where possible.** Domain types are written
  in the language they are for. A LLM author who needs to
  understand Series's eviction logic reads `series.ail`, not a
  Rust intercept registry. The implementation language matches
  the authoring language.
- **Composability.** Multiple library extensions can share one
  base extension. Series, Matrix, Hashmap, and future Vector all
  build on RawBuf. Each is a few hundred lines of AILang, not a
  new Rust intercept.
- **Cost localisation.** The Rust-intercept layer is small and
  centralised at the base. Domain growth (more library
  extensions) does not enlarge the Rust footprint.
- **Forward axis: SoA for records.** When the day comes that
  Series should hold records (e.g. `{price: Float, volume: Int}`),
  the change is internal to RawBuf — its intercept dispatches on
  whether the element type is primitive (flat buffer) or record
  (struct-of-arrays). Series's AILang code does not change.

## The plugin contract (consolidated)

Base extension:
- Kernel-tier module manifest.
- Rust codegen intercepts emitting LLVM IR for each operation.
- Optional C runtime — only when LLVM IR over existing runtime
  symbols cannot express the operation.

Library extension:
- Kernel-tier `.ail` module.
- Depends on one or more base extensions for its primitives.

Both register with the compiler via the kernel-tier
auto-injection mechanism described above. Core has no
per-extension code. The four mechanisms (type-scoped
namespacing, `new`, kernel-tier modules, `param-in`) are all
general — the same checker logic handles RawBuf, Series, and
any future extension.

The migration of existing hardcoded primitive-instance
intercepts (`try_emit_primitive_instance_body`) into the plugin
registry is triggered by the first real base extension shipping
(RawBuf). One registry mechanism for all intercepts, replacing
the current single-case hardcoded list.

CLI / discovery (`ail describe <T>`) consults the workspace
TypeDef registry and surfaces the `kernel: true` flag on the
home module — uniform for base and library extensions; the user
does not need to know which tier hosts the type.

## Series as the first library consumer (built on RawBuf)

### RawBuf — the base extension underneath

`RawBuf a` is a mutable, indexed, fixed-size buffer. Element
type `a` restricted to `{Int, Float, Bool}` for the MVP via
`param-in`. API:

- `(new RawBuf (con T) (size: Int)) : own (RawBuf T)` — allocates
  an uninitialised buffer.
- `RawBuf.get : forall a. (borrow (RawBuf a), Int) -> a` —
  indexed read. UB if `index >= RawBuf.size(b)`; caller checks
  bounds.
- `RawBuf.set : forall a. (own (RawBuf a), Int, a) -> own (RawBuf a)` —
  indexed write. Linear signature: caller hands over the buffer,
  callee returns it after mutation. Under uniqueness inference,
  the codegen rewrites this into in-place mutation; under
  shared ownership it falls back to a full buffer copy
  (Issue #22 territory).
- `RawBuf.size : forall a. (borrow (RawBuf a)) -> Int`.

Mutation discipline: ownership-mode-tracked, not effect-tracked.
The own/borrow signatures *are* the mutation contract. No
`RawBuf` effect is declared. This is consistent with AILang's
existing memory-model — RC + uniqueness as the canonical
mutation story.

Storage at the LLVM-IR level is type-specialised: the Rust
intercept registered for `RawBuf` dispatches on the element type
at each call site and emits the appropriate `getelementptr` /
`load` / `store` instructions over an opaque `ptr` allocated via
`@ailang_rc_alloc`. No new C code; the existing RC machinery
handles allocation, drop, and copy-on-share.

### Series — the library extension

`Series a` is a bounded ring buffer with financial-style indexing
(index 0 = newest). It is a plain AILang ADT in a kernel-tier
`.ail` module:

```
(module series (kernel)

  (data Series (vars a)
    (param-in (a Int Float Bool))
    (ctor S (own (RawBuf a))     ; the storage
            (con Int)              ; lookback
            (con Int)              ; head index
            (con Int)              ; current count
            (con Int)))            ; total pushes ever

  (fn new
    (type (fn-type
      (params (con Int))
      (ret (own (Series a)))))
    (params lookback)
    (body
      (term-ctor Series S
        (new RawBuf lookback) lookback 0 0 0)))

  (fn push
    (type (fn-type
      (params (own (Series a)) (con a))
      (ret (own (Series a)))))
    (params s v)
    (body
      (match s
        (case (pat-ctor S buf lookback head count total)
          (term-ctor Series S
            (app RawBuf.set buf head v)
            lookback
            (app % (app + head 1) lookback)
            (if (app < count lookback) (app + count 1) count)
            (app + total 1))))))

  (fn at
    (type (fn-type
      (params (borrow (Series a)) (con Int))
      (ret (con a))))
    (params s i)
    (body
      (match s
        (case (pat-ctor S buf lookback head count total)
          (app RawBuf.get buf
            (app % (app + (app - head 1) (app + (app * lookback 2) i)) lookback))))))

  (fn len
    (params s)
    (body (match s (case (pat-ctor S buf lookback head count total) count))))

  (fn total_count
    (params s)
    (body (match s (case (pat-ctor S buf lookback head count total) total)))))
```

Push semantics: take Series by `own`, return a new Series with
the buffer mutated (via `RawBuf.set`) and the bookkeeping fields
updated. The match-and-rebuild pattern is the explicit linear
state-threading style. Under uniqueness inference, both the
RawBuf mutation and the Series ADT rebuild become in-place; the
caller-side `(let s_new (app Series.push s v) ...)` pattern
optimises to single-allocation, two-counter-update cost. Under
shared ownership, push allocates a new Series wrapper (~40 bytes
RC overhead) per call — the performance fallback that Issue #22
will narrow.

No `Series` effect. The mutation is mode-tracked. Caller
signatures do *not* gain a `[Series]` or `[RawBuf]` effect entry;
they only need the linear threading discipline.

`Series.at` uses financial-style indexing (0 = newest):
`buf_index = (head - 1 - i + 2*lookback) mod lookback`. UB if
`i >= count`; the caller checks via `Series.len` or
`Series.total_count`. The `+ 2*lookback` is a non-negative
adjustment so the modulo arithmetic stays well-defined for
small `head` values.

### Forward axis: SoA for records (already accommodated)

Extending Series to hold records (`Series {price: Float, volume: Int}`)
is purely a RawBuf change. The RawBuf intercept gains a
dispatch path for record element types — primitive types stay
flat (AoS), record types switch to struct-of-arrays internally
with per-field parallel storage. `RawBuf.get` for a record-typed
buffer materialises the record on demand from the parallel
slots; `RawBuf.set` distributes the record's fields across the
parallel slots. Series's AILang code does not change; the API
surface is identical.

The Series milestone scope does *not* include record element
types; that is a future axis. But the architecture
accommodates it without revisiting Series itself.

## Feature-acceptance argument

Series's three-clause feature-acceptance check is in the spec; the
summary: LLM authors naturally reach for a bounded ring buffer
for streaming workloads (clause 1 — the SMA worked example
above is the evidence); the type structurally eliminates a class
of off-by-one and warmup-handling bugs (clause 2); the bounded
push-only mutation surface — gated by the `Series` effect — does
not reintroduce the iterated-mutable-state bug class that the
`mut`/`var`/`assign` removal addressed (clause 3 — see
`docs/specs/0052-kernel-extension-mechanics.md`'s clause-3
discussion).

The same three-clause check applies to *each future kernel
extension*. The whitepaper does not pre-justify Matrix or Decimal
or Time — each gets its own feature-acceptance gate when its spec
is written.

## Migration policy (pre-production frame)

AILang is pre-production. No external author depends on a
particular surface; no compiled artifact in the wild. The
implication for kernel-extensions: **when a new mechanism is
strictly better than an existing one for the use case it covers,
the existing one is retired or repositioned, not preserved
in parallel for compatibility.**

Concrete consequences for the prep milestone:

- Type-scoped namespacing makes `<module>.<TypeOrFn>` non-canonical
  for type-associated operations. Existing examples and fixtures
  using `std_maybe.from_maybe`, `std_pair.from_pair`, etc., are
  rewritten to type-scoped form. Hash pins are refreshed in
  lockstep — each test crate (`crates/ailang-core/tests/hash_pin.rs`,
  `crates/ailang-surface/tests/prelude_module_hash_pin.rs`, and
  any other hash-pin file the prep recon enumerates) is walked,
  per the "hash-pin blast-radius audit" practice.

- The existing `BareCrossModuleTypeRef` /
  `BadCrossModuleTypeRef` diagnostics are repurposed: what they
  considered "bad" before (a bare type ref to a foreign module's
  type) becomes "bad" still, but the *correct* form is now
  type-scoped (when feasible) rather than module-qualified.

- The prelude module gains `kernel: true`. The hardcoded
  prelude-name paths in `crates/ailang-surface/src/loader.rs:98-108`
  and `crates/ailang-core/src/workspace.rs:308-311, 467, 2655`
  migrate to the generic flag-driven mechanism. Consumer-
  observable behaviour for prelude is unchanged (the 12 free
  fns remain callable bare); the code path no longer hardcodes
  any module name as special.

- Codegen intercepts: the existing `try_emit_primitive_instance_body`
  hardcoded list is migrated into the plugin registry as part
  of the Series milestone — when there is the first real
  external consumer, the mechanism graduates from "hardcoded
  for one case" to "registry for many cases".

The reason migration cost is not a decision driver: there is
nothing to break externally, and rewrites inside the workspace
are cheap. Choosing the right design now is the priority; cost
of refactoring tests and fixtures is the project's own problem
and is amortised over zero external consumers.

## Coexistence with existing mechanisms

The mechanisms in this whitepaper interact with several
pre-existing parts of the language. Each interaction is named
here so the spec can quote the position:

- **Classes / method dispatch.** Orthogonal. Type-scoped
  namespacing accesses *top-level defs* in a type's home module;
  class-method dispatch (see [method dispatch](../contracts/0016-method-dispatch.md))
  resolves *class method calls* by type-driven instance lookup.
  Both can coexist on the same type — e.g. `Series` could
  someday have a `Show` instance (class method `show` dispatches
  via the instance) and a `Series.dump` free-fn def (accessed
  via type-scope). Both mechanisms remain.

- **Algebraic effects.** Reused unchanged. The `Series` effect
  is a new name in the flat effect-set; the `(do effect/op
  args)` syntax is the existing one (`io/print_str` is the
  precedent).

- **RC + uniqueness.** Series values are RC-managed like any
  heap value; drop is via a kernel-supplied `@ailang_series_drop`.
  Future Uniqueness inference (Gitea #22) can in principle
  recognise unique Series and elide RC ops on push — but the
  baseline does not depend on that.

- **Heap-Str ABI.** The Series runtime borrows the heap-Str
  RC-pointer-to-opaque-struct pattern (see
  [str-abi](../contracts/0011-str-abi.md)). Same pattern: opaque
  `ptr` from the compiler's view; dedicated `_new`, `_drop`,
  and operation symbols in C; RC header in the struct prefix.

- **Term::Ctor + pattern matching.** Unchanged. `term-ctor`
  and `pat-ctor` keep their roles for named-ctor ADTs.

- **`term-ctor` vs `new`.** Different constructs for different
  needs. The choice point for a hypothetical hybrid (e.g.
  "every named ctor implicitly defines a `new` fn") is
  out of scope for this whitepaper — the two coexist by
  design.

## Forward axes

This whitepaper does not commit to any future kernel extension.
The shape it enables — domain-specific types as plugins — is
intended to make a future Matrix, Decimal, Time, etc., a
self-contained crate-level proposal each gated by its own
feature-acceptance and brainstorm pass. The plugin contract
above is the only commitment.

## Cross-references

- **Feature-acceptance gate** applied to each kernel extension:
  [feature-acceptance](../contracts/0004-feature-acceptance.md).
- **Method dispatch** (orthogonal to type-scoped namespacing):
  [method-dispatch](../contracts/0016-method-dispatch.md).
- **Honesty rule** (this whitepaper's STATUS contract):
  [honesty-rule](../contracts/0007-honesty-rule.md).
- **Reference design** for Series storage layout:
  `/home/brummel/dev/RustAst/src/ast/rtl/series/` (external Rust
  project; not in this repo).
- **Per-milestone specs:** `docs/specs/0052-kernel-extension-mechanics.md`
  (prep milestone); future `docs/specs/YYYY-MM-DD-series.md`
  (Series milestone, written after prep closes).