spec: 0021 fan-in distinguishability

A construction-phase constraint (CompileError::IndistinguishableFanIn, raised in
inline_composite) makes a fan-in node whose sources have identical recursive
signatures illegal; the definition view then renders each fan-in input as the
shortest unique prefix of its source's signature (replacing positional #A/#B);
and the sma_cross fixture gains its missing fast/slow aliases.

Refs #44 — scope extended from render-only to a construction-phase invariant
during design (the positional #A was a symptom of unnamed, indistinguishable
fan-in sources, which is the actual fault to forbid).
This commit is contained in:
2026-06-08 14:08:59 +02:00
parent b73d7076c3
commit 14d91035ec
@@ -0,0 +1,351 @@
# Fan-in input distinguishability — Design Spec
**Date:** 2026-06-08
**Status:** Draft — awaiting user spec review
**Authors:** orchestrator + Claude
## Goal
A node with more than one input (a fan-in node) carries an ordered argument
list whose order is load-bearing for non-commutative nodes (`Sub`: a b). The
definition view currently labels those inputs with positional stubs derived
from the slot index — `[Sub(#A,#B)]` (`crates/aura-cli/src/graph.rs:182-183`).
`#A`/`#B` say nothing about *which producer* feeds each slot, and two distinct
`Sub` nodes in one graph render identically.
The root cause is upstream of the render. `sma_cross` (`main.rs:122`) wires two
`Sma` nodes into a `Sub` **without naming their length params** (empty alias
list, `main.rs:134`). Run `aura graph` and the signature reads `sma_cross() ->
(cross)` — no params at all — with two indistinguishable `[SMA]` boxes feeding
`[Sub(#A,#B)]`. The two SMAs are structurally and nominally identical at the
authoring surface; nothing says which is fast and which is slow. That is a
malformed composite, and the only reason it renders is that the render is a
best-effort total function. The contemporaneous `macd` fixture (`main.rs:191`),
authored after the graph view existed, does it right: aliased
`fast`/`slow`/`signal` lengths, a full `macd(fast:i64, slow:i64, signal:i64)`
signature.
This cycle makes that malformation **illegal by construction** and lets the
render rely on the guarantee:
1. **Construction constraint (aura-engine).** A composite with a fan-in node
whose two input slots are fed by sources with **identical signatures** is a
construction fault — a new `CompileError`, raised in the same per-composite
walk that already validates aliases, output ports, and role kinds.
2. **Source-derived render identifiers (aura-cli).** With distinguishability
guaranteed, the definition view replaces positional `#A/#B` with the
shortest abbreviation of each input's *producing node*`[Sub(#Sf,#Ss)]`,
`[Sub(#Ef,#Es)]` — answering "where does this input come from" from the label
alone.
3. **Fixture correction.** `sma_cross` gains `fast`/`slow` aliases, becoming a
well-formed composite that passes the new constraint.
### A node's signature (the recursive identity)
The signature is the authoring identity of a node, built depth-first:
```
signature(node) :=
typeInitial(node) -- 'E' for EMA, 'S' for SMA/Sub, …
++ [ initial(alias) for alias in node's param-aliases, in declared order ]
++ [ inputSig(slot) for slot in node's wired inputs, in slot order ]
inputSig(slot) :=
roleName(source) if the slot's source is an input role (named port — no descent)
compositeName(source) if the source is a nested composite (named — no descent)
signature(source) if the source is an interior leaf (recurse)
```
So `EMA(fast)` has signature `Ef`; `Sub(EMA(fast), EMA(slow))` has signature
`SEfEs`; a `Sma` with no alias fed by role `price` has signature `Sp`. The
recursion terminates because the dataflow is a DAG (C5); the only feedback path
is an explicit delay/state node, at which the descent stops (a state node is a
named type whose signature is its type + aliases, not its delayed input).
Two siblings **collide** when their full signatures are equal — which happens
only when their source sub-trees are structurally *and* nominally identical (or
are the very same node fed twice). The injected param *value* (length 2 vs 4)
is bound later and is not part of the signature; two alias-less SMAs on the
same input are identical regardless of the values they will receive.
The rendered identifier is the **shortest prefix** of a source's signature that
is unique among the consumer node's siblings (per-node-call namespace, not
composite-global; prefix-free among siblings). Most nodes resolve at the type +
first-alias initials (`Ef`, `Ss`) and never reach the recursive tail; nested
bare combinators reach into it (`SEf` vs `SEu`).
### Relationship to existing contracts
`param_space` documents (`blueprint.rs:176-177`) that "same-type siblings in
one composite share a name — uniqueness is at the slot". That stays true for
**param binding**: the injected vector is positional, one value per slot
(C12/C19/C23 untouched). This cycle adds a *separate* authoring well-formedness
rule that bites only at fan-in nodes: their sources must have distinct
signatures. C23 is not violated — the check runs during construction, before
the type-erased compilat; runtime names stay non-load-bearing debug symbols.
What changes is that a composite which was silently accepted (then rendered
ambiguously) is now rejected at compile. This is a deliberate contract
refinement, recorded in the ledger.
## Architecture
### Construction constraint (aura-engine)
`inline_composite` (`crates/aura-engine/src/blueprint.rs:344`) destructures a
composite into `{ nodes, edges, input_roles, param_aliases, output }` and runs
structural checks there (alias slot validity `:361-367`, output-port range,
role kinds). The new check joins them, before the interior is lowered:
- Enumerate each interior node's wired input slots from `edges` (`Edge.to ==
i`) and `input_roles` (`Role.targets` with `node == i`).
- For each node with **> 1** wired slot, compute each slot-source's full
signature (recursive, per the definition above).
- If two slots of one node carry equal signatures, raise the new
`CompileError`.
The check needs interior leaves' type labels and alias names at compile time —
`factory.label()` (the value-empty type name) and the `param_aliases` already
in hand — plus a recursive walk of interior `edges`/`input_roles` to build a
source's signature. No param *values* are required: the check is structural.
New variant on `CompileError` (`blueprint.rs:127`):
```rust
/// A fan-in node (>1 input) at interior index `node` has two input slots fed
/// by sources with identical signatures (type + alias names + recursive input
/// signatures), so the inputs cannot be told apart. Name a distinguishing
/// param (e.g. fast/slow), or the two sources are genuinely identical (a a).
IndistinguishableFanIn { node: usize },
```
### Render identifiers (aura-cli)
`leaf_label` (`crates/aura-cli/src/graph.rs:160`) keeps its slot collection;
the stub construction changes from "slot index → letter" to "slot's producer →
shortest-unique prefix of its signature":
- For each wired slot, resolve its producer and that producer's signature.
- The identifier is the shortest signature prefix unique among **this node's**
sibling inputs. A role passes its name through verbatim (`#price`). Each final
identifier is `#`-prefixed.
- Because the construction constraint guarantees distinct signatures at every
fan-in node, a unique prefix always exists for a *valid* blueprint
(lengthening terminates, at worst, at the full signature, which the
constraint guarantees distinct). No counter suffix.
- The render stays total: `aura graph` renders a blueprint *without* compiling
it, so it may be handed a malformed (constraint-violating) blueprint. There
the two signatures are equal and no prefix separates them; those slots fall
back to the positional letter (`#A`). A well-formed blueprint never reaches
this fallback — it is graceful degradation for a graph `compile()` would
reject, not a supported rendering.
No edge labels are introduced; `render_flat` keeps passing `None`
(`graph.rs:85`).
### Fixture correction (aura-cli)
`sma_cross` (`main.rs:122`) gains the two missing aliases, mirroring `macd`:
```rust
// params: was vec![]
vec![
ParamAlias { name: "fast".into(), node: 0, slot: 0 }, // fast SMA length
ParamAlias { name: "slow".into(), node: 1, slot: 0 }, // slow SMA length
],
```
## Concrete code shapes
### Worked user-facing example (the acceptance evidence)
`aura graph`, the `where:` definition of `sma_cross`. **Before** (today — the
defect): `sma_cross() -> (cross)`, two indistinguishable `[SMA]`, `[Sub(#A,#B)]`.
**After** the fixture correction:
```text
sma_cross(fast:i64, slow:i64) -> (cross):
[price]
┌────└────┐
↓ ↓
[SMA(fast)] [SMA(slow)]
└────┌────┘
[Sub(#Sf,#Ss)]
[cross]
```
The signature carries the lengths; each SMA shows its alias; the `Sub` reads
`[Sub(#Sf,#Ss)]` — "fast SMA minus slow SMA". The `macd` definition's two
`Sub`s, both `[Sub(#A,#B)]` today, become `[Sub(#Ef,#Es)]` (the line) and
`[Sub(#S,#Es)]` (the histogram) — distinct and self-describing.
### Nested bare combinators (the recursive case, now legal)
`Sub(Sub(EMA(fast),EMA(slow)), Sub(EMA(up),EMA(down)))` — two bare `Sub`s into
one `Sub`. The inner Subs have signatures `SEfEs` and `SEuEd`; distinct, so the
outer Sub is legal and renders by descending into the inputs just far enough:
```text
[Sub(#Ef,#Es)] [Sub(#Eu,#Ed)]
└──────┐ ┌──────┘
↓ ↓
[Sub(#SEf,#SEu)]
```
`#SEf` / `#SEu` = "the Sub whose first input is EMA(fast)" vs "…EMA(up)" — the
shortest prefixes that separate `SEfEs` from `SEuEd`. Deeper nesting grows the
identifier; an author who finds it unreadable *may* encapsulate the repeated
spread as a named composite, but is not forced to.
### The construction fault, shown
A composite shaped like the *old* `sma_cross` (two alias-less `Sma` on the same
role into a `Sub`) is rejected — both SMAs have signature `Sp`:
```rust
// two SMA leaves, no aliases, both fed by role `price`, both into one Sub:
// signature(node0) == signature(node1) == "Sp" (S + price)
// compiling a blueprint containing this composite:
// => Err(CompileError::IndistinguishableFanIn { node: 2 })
```
### Implementation shape (secondary)
`leaf_label` stub construction, today (`graph.rs:182-183`) maps slot index →
`#A/#B`; after, each wired slot resolves its producer's signature and renders
the shortest sibling-unique prefix (`#…`), with the positional letter kept only
as the malformed-graph fallback:
```rust
let stubs: Vec<String> = if slots.len() > 1 {
fan_in_identifiers(c, index, &slots) // signature prefixes, slot order
} else {
Vec::new()
};
```
The construction check inside `inline_composite`, beside the existing
validations (`blueprint.rs:361-409`):
```rust
// for every interior node with >1 wired input slot, its sources' signatures
// must be pairwise distinct, else IndistinguishableFanIn.
for node in 0..item_count {
let sigs = fan_in_source_signatures(node, &edges, &input_roles, &nodes);
if has_duplicate(&sigs) {
return Err(CompileError::IndistinguishableFanIn { node });
}
}
```
## Components
- **`CompileError::IndistinguishableFanIn { node }`** (new variant,
`blueprint.rs:127`).
- **construction check + `signature` walk** in `inline_composite`
(`blueprint.rs:344`): per-node fan-in distinguishability + the recursive
signature helper (type label + alias names + input signatures; role/composite
source stops the descent at the name). Reuses `factory.label()` and the
destructured `param_aliases`/`edges`/`input_roles`. Terminates on the DAG (C5);
descent stops at a state/delay node.
- **`leaf_label` / `fan_in_identifiers`** (`graph.rs:160`): shortest
sibling-unique signature prefix, role passthrough, positional fallback for the
malformed case. The wired-slot collection (today inline in `leaf_label`) is the
shared seam; it must not be duplicated.
- **`sma_cross` fixture** (`main.rs:122`): two aliases added.
- The signature notion is shared by the engine check and the CLI render (one is
equality-of-signatures, the other shortest-distinguishing-prefix-of-signature).
Whether the signature helper lives in aura-engine and is reused by aura-cli, or
is computed each side, is a plan-level call; its *definition* must be a single
source of truth.
## Data flow
Construction: `compile_with_params` → `lower_items` → `inline_composite` →
**new fan-in check** (before interior lowering) → on equal sibling signatures,
`Err(IndistinguishableFanIn)`; else continue as today.
Render: `render_definition` → `leaf_label(c, index, factory)` →
`fan_in_identifiers(c, index, slots)` → per slot: producer → signature →
shortest sibling-unique prefix → `#id`; structure only, no `eval`.
## Error handling
- **`IndistinguishableFanIn`** is the new, intended construction fault —
returned, never panicked; composes with the existing `CompileError` surface
(`PartialEq`, testable like `BadInteriorIndex`).
- **Render of a malformed blueprint**: stays total — equal-signature slots fall
back to positional letters; no panic, no dropped slot. A well-formed blueprint
never exercises this path.
- **Signature recursion**: bounded by the DAG (C5); descent stops at named
sources (roles, composites) and at state/delay nodes, so it cannot loop. A
prefix that reaches a source's full signature still resolves, since the
constraint guarantees the full signatures differ.
## Testing strategy
Engine (the central RED):
- **New constraint test** (`blueprint.rs` test module, beside
`BadInteriorIndex` / `RoleKindMismatch` ~`:782-817`): a composite with two
alias-less same-type leaves on the same role feeding one `Sub` compiles to
`Err(IndistinguishableFanIn { node: 2 })`.
- **Positive — aliased**: the same shape with distinct `fast`/`slow` aliases
compiles `Ok`.
- **Positive — nested bare combinators**: `Sub(Sub(a,b),Sub(c,d))` with
distinct interior leaves compiles `Ok` (distinct recursive signatures).
- **Negative — true tie**: `Sub(X, X)` (same node fed to both slots) compiles
to `Err(IndistinguishableFanIn)`.
Render (aura-cli):
- **`macd_blueprint_renders_a_nested_composite_definition`** (`main.rs:585`):
the single `[Sub(#A,#B)]` assert (`:598`) becomes `[Sub(#Ef,#Es)]` and
`[Sub(#S,#Es)]`.
- **Sample needle** (`main.rs:402`) and **`blueprint_view_golden`** (`:491`,
line ~522): re-captured for the corrected `sma_cross` — `sma_cross(fast:i64,
slow:i64) -> (cross)`, `[SMA(fast)]` / `[SMA(slow)]`, `[Sub(#Sf,#Ss)]`.
- **New focused render test**: per-node-call uniqueness (a producer abbreviates
independently at two consumers), role-name passthrough (`#price`), and the
recursive descent (`[Sub(#SEf,#SEu)]` for nested bare Subs).
- **`compiled_view_golden`** (`main.rs:533`): byte-stable — compilat labels via
`Box<dyn Node>::label()`, no fan-in stubs, no `#` identifier. C23 guard.
Fixture knock-on:
- Any test pinning `sma_cross`'s param-less form updates to the aliased form;
the sample `param_space` names gain `sma_cross.fast` / `sma_cross.slow`
(mirroring `macd_param_space_surfaces_the_three_named_aliases`, `main.rs:623`);
`sample_point` slot order/values unchanged (only names appear). The
grounding-check / planner enumerate the exact affected asserts.
## Acceptance criteria
Per aura's feature-acceptance: a strategy author cannot accidentally ship a
fan-in whose inputs are indistinguishable — the engine rejects it at
construction, where aura makes invariants structural; the graph view then
communicates input provenance correctly; no runtime/compilat behaviour changes
(C1/C12/C19/C23 intact).
- [ ] A composite with a fan-in node whose two sources have equal signatures
fails `compile()` with `IndistinguishableFanIn { node }`.
- [ ] Distinct param aliases, distinct nested combinators, and role/leaf mixes
compile `Ok`; `Sub(X,X)` (true tie) fails.
- [ ] `sma_cross` is corrected to declare `fast`/`slow` aliases; its signature
renders `sma_cross(fast:i64, slow:i64) -> (cross)`.
- [ ] A fan-in leaf renders each input as a `#`-prefixed shortest-unique
signature prefix (recursive into bare sources as needed), replacing
positional `#A/#B`; a role-fed input uses the verbatim role name.
- [ ] Abbreviations are scoped per-node-call; a valid blueprint never needs a
counter suffix.
- [ ] The render stays total on a malformed blueprint (positional fallback; no
panic).
- [ ] The MACD render shows its two `Sub`s as distinct labels; nested bare
combinators render their recursive identifiers.
- [ ] `compiled_view_golden` stays byte-stable (C23 guard); affected blueprint
goldens / needles / param_space asserts regenerated.
- [ ] The construction-constraint contract refinement is recorded in the design
ledger.