Aura/docs/design/INDEX.md

# aura design ledger — INDEX

The ledger records the load-bearing design contracts and their rationale. Each
contract states what it **guarantees**, what it **forbids**, and **why**. A
change that breaks a contract is a design decision (amend the contract here with
its new rationale), never a silent refactor.

Provenance: contracts C1–C18 were settled in the initial rough-sketch design
interview (2026-06-03), walking the design tree root-to-leaf (C16–C18 and the
C10 refinement to a broker-independent position table came in follow-up turns;
C10 was reframed in cycle 0007 to an exposure stream, then again 2026-06-23
(#117) to a **bias** stream with signal quality in **R**, the position table a
derived layer — see C10). C19–C22 were added as the
construction / World / playground layer; C23 and the C9/C19 compilation
refinements were settled 2026-06-05 for the **Construction-layer** milestone (the
blueprint→flat-graph reading of composites and graph optimisation). C24 (the
blueprint as a serializable, World-owned data value) was settled 2026-06-29,
resolving the long-deferred #109 fork in favour of **topology-as-data** (the
game-engine principle, C16): the engine owns topology as content it serializes /
loads / generates, not as Rust source baked into the binary.
The `CLAUDE.md` **Domain invariants** section is the compressed, always-loaded
summary of the
subset that agents must never violate; this file is the fuller form with
rationale.

Vocabulary: a *contract* is one ledger entry. A *cycle* is one pipeline round; a
*milestone* is a tracker container spanning many cycles (the first milestone is
the **walking skeleton**: ingest → one signal → exposure → sim-optimal broker →
synthetic pip-equity signal-quality metric — the *pre-reframe* substrate, now
reframed to bias → R-evaluator → R-expectancy, see C10).

---

## Foundation — what aura is

aura is a framework **and** a playground for traders. A human and (primarily)
LLMs author trading **nodes** directly in Rust; the engine backtests them
deterministically and massively in parallel, composes them fractally, validates
them (sweep / Monte-Carlo / walk-forward), and freezes a validated strategy into
a standalone bot with a broker connection.

The predecessor RustAst (`myc`) tried this as a custom DSL and failed: too slow,
too buggy, and LLMs author far better in Rust than in an unfamiliar DSL. aura
inverts it — engine in Rust, strategies in Rust — but keeps RustAst's *concepts*
(synchronous reactive streams, bounded-lookback series, run-counting, SoA).
RustAst is a conceptual reference, not a dependency. The one reused component is
`data-server` (the first data source).

What sets aura apart is **not** the single backtest — every quant system does
that — but the **World**: the meta-level where harnesses are dynamically
constructed and *families* of them orchestrated and explored (walk-forward,
sweeps, Monte-Carlo, comparison). The deterministic single-harness engine is the
*substrate*; the World is the *product* (C20–C22).

---

## External components

Two sibling projects live outside this repo and are named throughout the
contracts. Their concrete location is recorded here so the repo is the single
source of truth — not session memory.

- **`data-server`** — `~/dev/libs/data-server`, Gitea `Brummel/data-server`
  (`http://192.168.178.103:3000/Brummel/data-server.git`). aura's **first data
  source** and the one reused component (Foundation; C3, C11, C12). A standalone
  leaf crate (deps: `chrono`, `regex`, `zip`) that loads Pepperstone **M1/tick**
  binary files and shares them lock-free as `Arc<[T]>` chunks
  (`CHUNK_SIZE = 1024`) via `SymbolChunkIter::next_chunk` (an inherent method
  driven by `while let`, not the `Iterator` trait); `stream_*_windowed(from_ms,
  to_ms)` provides C12's data-window with inclusive Unix-ms bounds. Records are
  **AoS** (`M1Parsed`/`TickParsed`, `time_ms: i64` Unix-ms). The ingestion
  boundary (C3) therefore **transposes** these AoS records into aura's SoA
  columns (C7) and normalizes `time_ms` → canonical epoch-ns at that one
  boundary. Pulled in as a cargo git dependency by the **`aura-ingest`** crate
  (cycle 0011) — the **data-source ingestion edge**, where the `data-server`
  external tree (and its transitive `chrono`/`regex`/`zip`) enters the workspace.
  Under the amended C16 **per-case dependency policy** (cycle 0029), `aura-ingest`
  is no longer a zero-dep *firewall* for the rest: the engine crates link
  standard vetted crates where they earn their place (e.g. `serde`/`serde_json`
  for the run registry, cycle 0029), with particular scrutiny for the frozen
  deploy artifact (C13). Consequence: `cargo build/test --workspace` resolves
  from a populated cargo cache (a Gitea fetch for `data-server` + crates.io for
  the standard crates) rather than fully offline.
- **RustAst (`myc`)** — `~/dev/RustAst`, Gitea `Brummel/RustAst`
  (`http://192.168.178.103:3000/Brummel/RustAst.git`). The predecessor DSL attempt
  (Foundation): **a conceptual reference, never a dependency** — its DSL authoring
  surface and `Value`/HM-inference machinery are exactly what aura rejects (C17).
  But its **`src/ast/rtl/` layer is a working reference implementation of the very
  streaming substrate aura rebuilds**, and is worth reading before authoring
  `aura-core` — these are not just concepts, they exist as code:
  - `rtl/series/data.rs` — `RingBuffer<T>` with financial-style indexing (index 0
    = newest), `total_count` (the **run-count** of C5) and `lookback_limit` (C8's
    pre-sized window); the `ScalarValue` marker trait ("flat scalars only, no
    String/Record" = C7's closed scalar set); `ScalarSeries<f64|i64|bool>` and the
    **SoA** `RecordSeries` for composites (C7's "OHLCV = a bundle of base columns").
  - `rtl/series/mod.rs` — `create_typed_series`, dispatching element type → storage
    backend (the type-specialized **factory** of C19).
  - `rtl/streams/mod.rs` — `Signal { cycle_id, value }` (C4's cycle clock) and the
    `Stream` / `Observer` / `ObservableStream` push traits (the reactive model of
    C4/C5).
  - `rtl/streams/register.rs` — the RTL **"register" / delay node** (the one
    explicit feedback path of C5/C9) plus a seeded, reproducible OHLC generator
    (C12's seed-as-input).

  aura reimplements these natively and **sharpens** them: types are monomorphized
  and edges type-erased to the four scalar kinds with direct dispatch (C7) instead
  of carrying boxed `Value`s, and input history is shared zero-copy as `Arc<[T]>`
  (C12) instead of a `VecDeque<Value>`. RustAst shows the *shape*; aura makes it
  fast and deterministic.

---

## Contracts

### C1 — Determinism and disjoint parallelism
**Guarantee.** A backtest is a deterministic, synchronous, non-concurrent event
loop that reaches a unique state after each input tick. Same input (incl. seed)
→ bit-identical run. Two backtests are fully disjoint and run concurrently
without locking.
**Scope (ratified, fieldtest 0078).** The bit-identity is *per run* — one backtest
of given (inputs, seed) reproduces byte-for-byte. A *derived metric* recomputed for
the same params by two *different command paths* (e.g. a swept member's `sqn` vs the
same cell re-run under `aura generalize`) may differ by floating-point reassociation
(≤1 ULP), since the two paths accumulate the same logical reduction in a different
operation order (IEEE-754 non-associativity). This is not a C1 violation: C1 governs
the determinism of a single run, not the cross-command bit-identity of a re-derived
statistic.
**Forbids.** Concurrency *within* a single sim; any nondeterministic input that
is not captured as an explicit input (see C11, C12).
**Why.** Real money rides on backtest results; reproducibility and an audit
trail are non-negotiable. Speed comes from parallelism *across* sims, which
disjointness makes lock-free.

### C2 — Causality / no look-ahead
**Guarantee.** A node sees only the past. Input history is a read-only window
that ends at the current cursor; a resampler emits a bar only once it is
complete.
**Forbids.** Any node access to data with `timestamp > now`; emitting a partial
/ still-forming bar.
**Why.** Look-ahead is the cardinal backtester bug — a fast backtester that
leaks the future is worse than none. Making the future *physically absent* from
what a node receives beats merely discouraging it.

### C3 — One merge, at ingestion only
**Guarantee.** Heterogeneous timestamped sources are k-way-merged by timestamp
into one chronological cycle stream at the ingestion boundary; source-native
time units (e.g. data-server's Unix-`time_ms`) are normalized there to the
canonical epoch-ns `timestamp` of C7.
**Forbids.** Any merge / as-of join *inside* the graph.
**Why.** A single ordered timeline is the mechanism that makes heterogeneous-rate
sources (news daily-bias + M5 + ticks) causally combinable without leaking the
future. Keeping the merge at one boundary keeps the graph semantics simple.

### C4 — Cycle granularity
**Guarantee.** The clock is data-driven: one input record = one cycle, advanced
in global timestamp order, with a monotonic `cycle_id`. Ties (same timestamp,
multiple sources) break by source declaration order.
**Forbids.** A fixed time-grid clock; nondeterministic tie ordering.
**Why.** The market *is* an irregular event sequence; a grid is arbitrary and
either wastes empty cycles or clumps ticks. Backtest and live differ only in the
origin of records, not the cycle semantics. Tie determinism preserves C1.

### C5 — Freshness-gated recompute and sample-and-hold
**Guarantee.** The `cycle_id` advances everywhere (a cheap counter), but a node
re-evaluates only when ≥1 of its own inputs is fresh this cycle (detected by
run-count); otherwise it holds its last output. Stale inputs contribute their
last (held) value.
**Forbids.** Recomputing every node every cycle ("push all" is true for the
*clock*, not for *recompute"); treating a held value as missing.
**Why.** Total recompute does not scale to many sparse high-frequency sources;
freshness-gating is the performance discipline that keeps the synchronous model
fast.

### C6 — Firing policy A and B, per input group
**Guarantee.** A node declares, per input group, one of two firing policies:
**A** fire-on-any-fresh + hold (latest / as-of join — e.g. tick × held
daily-bias); **B** all-fresh barrier (synchronizing join — e.g. O/H/L/C from
four separate 15m sources: the candle is complete only when all four are fresh).
A single node may mix an A input and a B group.
**Forbids.** A single global firing mode; forcing per-node-only granularity.
**Why.** Both are genuinely needed; RustAst implemented only B. Per-input-group
granularity is required by the OHLC-plus-bias case where one node needs both.
**Realization (cycle 0004).** Firing is tagged per input — `Firing::{Any,
Barrier(u8)}` on `InputSpec` — and inputs sharing a `Barrier` id form a group; a
mode-A input is its own trivial group, so "per input group" and the per-input tag
coincide. The barrier's synchronization token is the cycle **timestamp**, not the
`cycle_id`: under C4 four same-timestamp sources are four distinct cycles, so
RustAst's `cycle_id`-equality barrier could never fire across them. A group fires
when every member's last push carries the current cycle's timestamp, guarded by
"≥1 member fresh this cycle" (so a group completed earlier does not re-fire). This
fires both the multi-source bar and the within-source diamond rejoin (every push
in a cycle carries that cycle's timestamp).

### C7 — Four scalar base types, streamed as SoA
**Guarantee.** Only `i64`, `f64`, `bool`, `timestamp` (newtype over i64,
epoch-ns UTC) are streamed, as columnar Structure-of-Arrays. Composite streams
(OHLCV) are bundles of base columns — this is the **node-output model** too: a
node emits a record of 1..K base columns (C8), each forwarded field-wise to a
consumer slot; the bundle is structural, never a fifth scalar type. Edges are
type-erased to these four kinds;
the type check is paid once at wiring/sim-start, then the topology is frozen per
sim → direct dispatch, no per-event allocation.
**Forbids.** Streaming non-scalars (String, Records, tables, calendars) — those
live as metadata beside the hot path; `dyn Any` payloads; per-event heap
allocation; topology mutation mid-sim.
**Why.** Maximal streaming performance (SIMD/cache) needs a tiny closed scalar
set and SoA. The open set is composites (schemas of columns), not scalar types.
Type-erasure at the edge is also forced by the cdylib boundary (C13).
**Realization (`Cell` carrier split, 2026-06).** The streamed value is now split
into a tag-free 64-bit word and its kind. `Cell` (`crates/aura-core/src/cell.rs`)
is a type-erased `u64`: constructed per base type (`from_i64/f64/bool/ts`) and
read only by naming the type at the call site (`i64()/f64()/bool()/ts()` —
branch-free bit-casts). The kind is therefore resolved once at the boundary and
the value itself carries no tag — C7's "the type lives at the column/edge, not in
the value" made explicit on the single-value carrier (and a single 8-byte word
vs. the 16-byte tagged enum). `Scalar` becomes `{ kind: ScalarKind, cell: Cell }`
— the self-describing form for the *dynamic* boundaries (builder binding, serde,
rendering) — with `debug_assert`-guarded native accessors (caller asserts the
kind; free in release) and a hand-written **value** `PartialEq` that preserves
the former enum's IEEE-754 semantics (`NaN != NaN`, `+0.0 == -0.0`), pinned by
the `scalar_eq_is_value_not_bitwise` fixture. Behaviour-preserving.

**Realization (`Cell` becomes the hot-path carrier, 2026-06, #74).** The carrier
swap deferred above has landed: `Node::eval` now returns `Option<&[Cell]>` and
every node out-buffer is `[Cell; N]`, so the inter-node forward carries tag-free
8-byte words. The kind lives only at the schema/column: the harness forwards each
field via the new branch-free `AnyColumn::push_cell` (infallible — the edge
kind match is verified once at bootstrap, the surviving guard
`bootstrap_rejects_*_kind_mismatch`). `Scalar` remains on the self-describing
dynamic boundaries: the param plane (`build`/`bind`/`compile_with_params`/sweep
points/`RunManifest`), `AnyColumn::get` (the type-erased read for sinks/serde),
and source ingestion (`Source::next`, the heterogeneous C3 merge). The removed
per-value runtime kind check on node output is the same authoring-bug class C8
already leaves to a `debug_assert` (output width); node-output-kind correctness
is the node's declared `FieldSpec` contract, caught by each node's own test.
Behaviour-preserving (C1).

### C8 — The node contract
**Guarantee.** A node has a **signature** — its `NodeSchema`: each input's scalar
type and firing group, its output record, **and the node's own tunable parameters —
typed, with ranges**, which aggregate into the blueprint's param-space the optimizer
sweeps (C12/C19/C20). The signature is **declared pre-build on the value-empty
recipe** (C19), so the whole interface is legible without building (cycle 0024).
The built node implements `lookbacks()` — the per-input buffer depth, the one
quantity that may depend on an injected param (e.g. an `Sma`'s window = its
`length`), read once at bootstrap to size the windows — and `eval(ctx) ->
Option<&[Cell]>`. The
engine provides read-only, zero-copy windows into each input's SoA ring buffer
(`ctx.f64_in(x)[k]`, sized at wiring); a node may *additionally* keep its own
mutable series for derived/intermediate state. `None`/Void return = filter /
not-yet-warmed-up. A node is a **producer, a consumer, or both**: a
producer/transformer exposes **one output port**, whose payload is a **record of
1..K base-scalar columns** (a scalar is the degenerate K=1 record; an `eval`
returns a borrowed row, one value per column); a **pure consumer (sink)** — chart,
equity, logger — has **no** output. Sources are pure producers; sinks are pure
consumers.
**Forbids.** A node sizing/growing its input lookback at runtime; more than one
output **port** per node; a fifth scalar type or a heterogeneous output payload
(a record is a bundle of base columns, C7); copy-on-read of input history.
**Why.** Engine-provided windows mean LLM-authored code cannot mis-manage
lookback bookkeeping, and history passes through zero-copy. Fixed, pre-sized
buffers suit deterministic, pre-dimensioned sims (no realloc in the hot loop).
**Realization (cycle 0005).** `NodeSchema.output` is a `Vec<FieldSpec>` (named base
columns; length 1 = scalar). Binding is **field-wise only**: `Edge::from_field`
selects one producer column per edge; consuming a whole record is N edges (no
"bind whole record" mechanism). The K fields of one record are **co-fresh by
construction** (one `eval`, one timestamp), so C6 is untouched. `eval` returns
`Option<&[Cell]>` — a borrowed row into a node-owned buffer — so the forward
path allocates nothing per cycle (C7) (the carrier is now a tag-free `Cell` — see
the C7 carrier note).
**Realization (cycle 0006).** The pure-consumer (sink) half of this contract is
now realized at the substrate: **recording is a node role, not a type.** A
recording node reads its typed input windows + `ctx.now()` in `eval` and pushes
the record to a destination it holds as a field (a channel, a chart handle) — an
**out-of-graph side effect**. There is no `Sink` type, trait, or engine flag: a
node that only records returns `None` (pure consumer), and a node may record
**and** return an output the engine forwards in the same `eval` (the "both"
case). **Encoding & return contract.** A pure consumer declares `output: vec![]`
— the empty record *is* the sink declaration; there is no separate type, trait,
or marker. Its `eval` returns `None` or a zero-width `Some(&[])`, and the run
loop debug-asserts the returned row's width equals the declared output width
(`row.len() == schema.output.len()`). Field-wise wiring resolves
`Edge::from_field` against the producer's `output` at bootstrap, so no edge can
bind a field of a zero-output node — it fails with `BadIndex` — making a sink
structurally unwireable as an in-graph producer; its only output is the
out-of-graph side effect. In-graph routing stays engine-owned data (the edge table); the escape out
of the graph is the node's own side effect — and that boundary is the
determinism / graph-as-data boundary (C1/C7).

**Refinement (cycle 0070 — end-of-stream `finalize` lifecycle, 2026-06-25).** The
node contract gains a second lifecycle phase beside `eval`: `finalize(&mut self)`,
a **default-no-op** trait method the engine calls **once per node, in topological
order, after the source loop drains** (the `Harness::run` epilogue). It lets a sink
**fold online** — accumulate per `eval` into O(trades)/O(1) owned state and flush
one compact summary at stream end — instead of pushing a record every fired cycle
into an unbounded channel that buffers O(cycles) rows until the run ends (the
retention BLOCKER #138 hit: a full-history Stage-1 sweep held ~2 GiB/member over
~5.5M one-minute bars). C1/C7/C8 are preserved: `finalize` runs once *after* the
deterministic event loop, adding no within-sim concurrency (C1); a folding sink
holds only owned accumulator state, no interior mutability (C7); it still declares
`output: vec![]` and its flush is the same out-of-graph side effect (C8) — one
summary row, not a per-cycle stream. Realized by `GatedRecorder` (emits only gated
rows + the genuine final row) and `SeriesReducer` (folds one column, emits one
summary row), aura-std siblings of the per-cycle `Recorder`, which survives for the
live / `--trace` / test-tap path. The recording sink's *own* per-cycle allocation
(the `Recorder`→`Probe` rename + its accumulate-vs-stream choice, #77) stays open;
`finalize` is the "non-channel exit from the graph" that issue's
accumulate-then-read option named as missing.

**Refinement (Construction-layer milestone — render labels, 2026-06-05).** A node
additionally exposes `label() -> String`, a **single-line, non-load-bearing**
render symbol: a default trait method the run loop never calls (wiring is by
index, C23). Overrides carry the node's identifying params (`SMA(2)` vs `SMA(4)`)
so a graph render (C9 graph-as-data, #13) disambiguates identical node types and
surfaces a mis-wiring. Like `FieldSpec.name`, it is an informative debug symbol,
not part of the C8 dataflow contract — adding it changes no run behaviour.

**Realization (cycle 0015 — param declaration).** The tunable-parameter half of
this contract is now realized: a node declares its knobs in `schema()` as
`params: Vec<ParamSpec>` (`ParamSpec { name: String, kind: ScalarKind }`), and
`Composite::param_space()` (cycle 0024; was `Blueprint::param_space()`) aggregates
them into one flat, path-qualified list — a
read-only projection of the graph-as-data (C9), mirroring the inline order
(C19/C23) so a param's slot matches the later flat-node order. Two refinements to
the guarantee's "typed, with ranges": (1) the declaration carries `name` + `kind`
only — the **search-range is the run's, not the node's** (which subset / grid a
sweep covers is an experiment axis, #32/C20; the node declares the knob's existence
and type, never its search interval). (2) Identity is **positional** (the slot,
C23 "by index, not name"); the path-qualified `name` is a non-load-bearing debug
symbol (like `FieldSpec.name`) — uniqueness is at the slot. **Refinement (cycle
0032):** identity in the **flat graph** stays positional (C23, unchanged), but the
`param_space()` **name projection** — the authoring / by-name address space (the
surface a sweep axis or single-run binding addresses) — must be **injective** for a
blueprint to compile (a duplicated path is unaddressable; see C9). Two layers:
positional wiring below (the flat graph), an injective name address space above (the
authoring boundary). So same-type siblings in one composite no longer silently
share a name — they must be `.named(...)` apart, which is also what disambiguates a
same-type fan-in (C9). A vector knob (`LinComb.weights`) expands to `N` flat
`weights[i]` entries, `N` topology-fixed (C19). Permitted kinds are `i64`/`f64`/
`bool` (a `timestamp` knob is a structural axis, C20, never a numeric sweep param).
Binding a value to a slot landed in cycle 0016 (#31, see C19/C23); enumerating a
sweep (#32) is the deferred next layer. The 0016 binding also moved the param
declaration's *authoring* home into the value-empty recipe, whose `params()` is read
pre-build. In 0016–0023 the built node's `schema().params` still reported the same
slots, kept in lockstep by a per-node test (a duplication filed as debt, #36); **cycle
0024 dissolved this** — the signature is declared *once* on the recipe and the built
node no longer carries `schema()` (#36 closed). See the C8 cycle-0024 realization.

**Realization (cycle 0024 — the signature lives in the blueprint, #43/#36).** The
node's whole declared interface — its `NodeSchema` (input scalar types + firing,
output record, params) — now lives **once**, on the value-empty recipe
`PrimitiveBuilder` (ex-`LeafFactory`), read pre-build by `param_space()`, the
compile-time structural validation, and the render. This closes two debts: **#43**
(a value-empty recipe used to declare *no* input/output interface pre-build — only
params) and **#36** (params declared twice, recipe vs built node, kept in lockstep
by a per-node test — those 8 tests are deleted, the duplication structurally gone).
The split that makes this work: a node's signature is **fully static** per blueprint
(input kinds/firing, output fields, params — verified across the roster; a variable-
arity node like `LinComb` takes its arity as a *recipe argument*, not an injected
param), so it can be declared without building; the **one** param-dependent quantity,
an input's buffer **lookback** depth, is no longer in the signature but answered by
`Node::lookbacks() -> Vec<usize>`, read only by bootstrap to size the windows.
`Node::schema()` is therefore **removed** — the signature is pre-build data, not a
built-node method. `BlueprintNode::signature()` answers uniformly for both arms (a
primitive returns its recipe's schema; a composite *derives* it from the interior:
role kinds in, re-exported field kinds out, aggregated params), so "every node has a
signature in the blueprint" holds for composites too.

**Realization (cycle 0027 — name input ports, refs #21/#51).** `PortSpec` now
carries a `name: String` (it drops `Copy`, like `ParamSpec`), so an input port is
named just as `FieldSpec.name` (output) and `ParamSpec.name` (param) already are.
The name is **non-load-bearing** (C23): wiring stays positional by slot, bootstrap
and the run loop never read it; it exists for tracing / graph rendering (#13). Leaf
primitives declare their slot names (`SimBroker`'s `exposure`/`price`, etc.);
`derive_signature` carries a composite's `Role.name` into the derived input port,
so the graph model (`model_to_json`) is homogeneously named across inputs, outputs,
and params and across both graph levels. This does **not** close #21 (a swapped
same-kind wiring is still only kind-checked) — it makes those slots
self-documenting; a name-consuming validation is its own future cycle.

**Realization (cycle 0040 — wiring totality: every input slot connected exactly
once, #65).** The node contract's "the engine provides a window into each input"
presupposes each input is actually fed; until now an unwired interior input slot was
accepted and bootstrapped to a silent empty column (`harness.rs`), and two producers
into one slot — ill-formed, since a slot holds one column — was likewise
uncompiled-against. Cycle 0040 makes a valid graph **total and single-valued** over
its interior input slots: `check_ports_connected` (in `validate_wiring`, run at every
nesting level) requires every interior node's every declared input slot to be covered
by **exactly one** wiring act — one interior `Edge { to, slot }` or one role
`Target { node, slot }`, edges and role targets counted uniformly. Zero coverage is
`CompileError::UnconnectedPort`, more than one is `CompileError::DoubleWiredPort`. A
composite's **own** input roles (`source: None`) are coverage *providers* — the
wired-by-enclosing boundary, the root case already guarded by `UnboundRootRole` —
never consumers, so only interior input slots are subject to the rule. There is **no
optional-input concept**: every declared input port is required (no shipped node runs
meaningfully without one; an unwarmed mode-A input is *wired-but-not-yet-valued*, not
unwired). The check is **index-based and name-free** — it touches no name machinery
and emits nothing into the flat graph, so C23 is untouched (it proves the existing
raw-index wiring is total and single-valued). Inherited identically by the raw
`Composite::new` path and the `GraphBuilder::build()` path (both compile via
`compile_with_params`).

### C9 — Fractal, acyclic composition
**Guarantee.** A composite is itself a `Node` that wires a sub-graph and exposes
one output; signal, combined signal, and (with execution) strategy are all the
same abstraction, nestable arbitrarily. The dataflow graph is a DAG; the only
feedback path is an explicit delay/state node (the RTL "register"). Wiring is
written in Rust (builder API); the built graph is introspectable runtime data.
**Forbids.** Implicit dataflow cycles (combinational loops); special-casing
"signal-of-signals" as separate mechanics.
**Why.** Self-application of one contract gives unlimited composition with no
adapter zoo. Acyclicity keeps the synchronous reactive model well-defined;
forcing feedback through a visible delay node keeps the per-cycle determinism
intact and the one legitimate feedback path explicit. Graph-as-data enables
visualization, freezing, and re-parameterization for sweeps.
**Refinement (Construction-layer milestone, 2026-06-05).** "A composite is itself
a `Node`" is an **authoring-level** identity: a composite declares the same
interface (typed inputs + ≤1 output, C8) and is wireable wherever a node is, but it
is **not** a runtime object. The bootstrap **compiles it away by inlining** its
sub-graph into the one flat instance (C19/C23): the composite *boundary* dissolves
into the raw index wiring the run loop already consumes, and names stay
**non-load-bearing** (informative debug symbols only, as `FieldSpec.name` already
is). So "nestable arbitrarily" and "graph-as-data" hold at the **blueprint
(source)** level; the running graph is the flat, index-wired **`FlatGraph`**. The earlier reading — *a composite survives as a `Box<dyn Node>`
driving a nested sub-engine* — is **explicitly rejected**: it would keep the
interior opaque to the cross-graph optimiser (C23) and add a runtime sub-loop the
flat model does not need. Inlining is what makes the composite boundary free.
**Realization (cycle 0018 — composite multi-output record, #40).** "Exposes one
output" is one output **port** carrying a **record of 1..K re-exported fields**, not
one field: `Composite.output` is a `Vec<OutField { node, field, name }>` (was a
single `OutPort`). Each entry is a **named projection** of one interior
`(node, output-field)`; a consumer selects which re-exported field it reads via the
same `Edge::from_field` that already binds leaf record columns (C8 realization,
cycle 0005). This is the **same arity** C8 already grants a leaf (OHLCV = one port,
5 columns): a multi-line indicator (MACD = macd/signal/histogram) is **one** record
of K fields, **one** port, **one** row per `eval` — C8/C7/C4 untouched, a boundary
completion, not a contract change. A strategy composite is simply the K=1 case
(one bias field, C10). The re-export **names** are **non-load-bearing** (C23):
they live at the blueprint boundary and in render (cycle 0022/#46 folds each onto
its producing node as a `name := …` binding; originally `[out:<name>]` markers, #13)
but are dropped at lowering — `ItemLowering::Composite.output` is `Vec<(usize,
usize)>`, raw index pairs only, so the flat graph is name-free (verified: the
compiled-view render stayed bit-identical across this change).
**Realization (cycle 0019 — name the composite boundary, #41; param-overlay retired,
cycle 0031).** The named-projection shape covers the surviving boundary edge-kinds:
`input_roles` is a `Vec<Role { name, targets }>` (was a bare `Vec<Vec<Target>>`),
alongside `output: Vec<OutField>`. Each is an ordered, positionally-indexed
**named projection** of interior handles. **Param projection is no longer a
composite overlay** (the index-addressed `ParamAlias` was retired in cycle 0031):
a node's surface param name flows from its own **instance name** — every node
carries a name (default = its lowercased type label, override via `.named()`), and
`param_space()` is uniformly `<node>.<param>` at every level including the root. A
same-type fan-in is distinguished by naming the colliding legs, the same single act
that qualifies their param paths. Like the output and role names, **node names are
non-load-bearing** (C23): they live at the blueprint boundary and in render but are
dropped at lowering — the flat graph is wired by raw index. The full composite
boundary signature (named inputs, multi-outputs) and the per-node param path are
legible without changing the flat graph.
**Refinement (param-namespace injectivity, cycle 0032; supersedes the fan-in
distinguishability check).** A blueprint compiles only if its `param_space()` name
projection is **injective** — every path-qualified knob name is unique. A duplicated
path is a knob no binding can select alone (it has no distinct by-name address,
C12/C19), so it is a `CompileError::DuplicateParamPath` carrying the offending path;
the cure is to give the colliding same-type sibling nodes distinct names with
`.named(...)`, the same single act that qualifies their param paths. The
param-bearing indistinguishable fan-in is *one instance* of a duplicated path (two
default-named same-type legs share a leaf path). `signature_of`, `leaf_has_param`,
and the fan-in-specific check (`check_fan_in_distinguishability` /
`check_composite_fan_in`) are **retired** (cycle 0032), replaced by one structural
`check_param_namespace_injective` over the `param_space()` names, run before name
resolution in `compile_with_params` and both binders — so the canonical by-name
author sees the structural cure rather than a downstream `AmbiguousKnob` symptom
(that `BindError` arm is retired too: an injective space can never multi-match).
Paramless interchangeable same-name sources stay legal (no path, no duplicate).
The old signature-collision predicate had extra breadth — it also rejected an
**asymmetric param/paramless** collision and a **role-vs-leg** collision, *neither*
a path duplicate (each colliding configuration keeps a unique param path, or none).
That breadth guarded **render identity**, dead since the renderer was retired in
0026; both extra rejections are **intentionally dropped**. A future node-/wiring-name
distinguishability check, if ever wanted (e.g. for the WASM graph view's #21
thread), is decoupled from param-space injectivity. Construction-phase only; the
flat graph stays name-free (C23).
**Refinement (2026-06-29 — graph-as-data round-trips both ways, C24).** "The built
graph is introspectable runtime data" is now contract-level **bidirectional**: a
blueprint is not only *emitted* as data (`model_to_json`, the render half) but is
itself a **serializable data value with a load path** (data → blueprint →
`FlatGraph`), so topology is a value the World generates, stores, and reproduces —
see C24. The Rust builder API stays the primary *human / LLM* authoring surface
(C17); the data form is the *machine* surface the World owns. (Load path **realised**
cycle 0087 / #155, `d5602ec`: `aura-engine::blueprint_from_json`.)

### C10 — Strategy output is a bias stream; signal quality is measured in R; cost is a composable downstream graph (gross R → net R); money is decoupled to the live deploy edge
**Guarantee.** A strategy's primary, backtestable output is **not** an equity
curve, **nor a position-event table**, **nor a position size**, but a **bias
stream**: the DAG expresses exactly one *state* at time t (C8 — a node emits at
most one record per `eval`), so a strategy emits one **signed, bounded bias**
`f64 ∈ [-1, +1]` per cycle (per instrument) — the **sign is direction, the
magnitude is conviction**, and conviction is optional (a bare ±1 / 0 is the modal
case). Bias is **unsized**: position sizing and the protective stop do **not**
live in the strategy. The chain is `signals (scores) → decision node → bias
stream`.

**Risk-based execution is a decoupled downstream layer; in research it is Stop +
position-management in R, no Sizer.** Turning a bias into a tracked trade is the
job of a downstream **execution** chain, never the strategy's. The **stop-rule**
sets a protective stop, which **defines the risk unit R** (1R = the loss taken if
stopped). In the research loop the executor is **stop-rule → position-management**,
operating **directly in R**, with the **Veto** an optional documented pre-trade-gate
seam (a pass-through identity DCE'd away under C19/C23 when absent): there is **no
Sizer**. Sizing in *currency* (`size` / `volume`) is a **deploy** concept. C8's
wiring totality (cycle-0040 `check_ports_connected`: no optional-input concept —
every declared port is covered by exactly one wiring act) forbids a *dangling*
`size` port, so the resolution is concrete: research `PositionManagement` either
**drops its `size` input port** from its node signature, or has that port **driven
by a constant unit node** (the flat-1R degenerate) — never an unwired "vestigial"
port; the record's `size` field is held at unit and carries no research
information, and the position-event table's `volume` column likewise. Consequently
the **position-event table is demoted to a deploy / reconciliation artifact** (real
volume lives there), no longer a research artifact and no longer "fed to a broker".
The three-way decoupling of **direction (bias) / sizing / fill** stands as a
*structure*; sizing and fill are simply pushed entirely to the deploy edge, out of
the research loop.

**Signal quality is measured in R — gross R and net R.** A downstream
**R-evaluator** consumes the executor run and integrates the per-trade R-outcomes
into an **R-expectancy** / R-curve: the account- and instrument-agnostic yardstick
for *"how much R out per 1R risked?"*. R, not pips, is the unit (pips are not
risk-normalized). Two readings of the same unit: **gross R** (signal only) vs **net
R** (after the cost model), with `net R = gross R − cost-in-R`. The headline
artifact is the **net-R equity curve** — the cost-drag drawn onto the R curve,
recorded through a named **`net_r_equity`** tap/sink (sibling of the existing
`r_equity` tap; a sink is the only thing the registry can display, C8/C18/C22). No
new unit is invented: it is R, gross and net, **continuous with the existing
`net_expectancy_r`**. A bare **gross-R run with no cost model attached is valid**:
the cost layer is optional and additively composed-on (the zero-cost baseline is
the "default simple" floor).

**Cost is a COST MODEL — a composable C9 graph of cost nodes, in R, that
approximates (never claims) realism.** The realistic broker is *retired* (see
Forbids / the 2026-06-28 reframe): real friction — slippage (live
liquidity / order-size / volatility at fill), swaps (broker-set, time-varying),
even recorded feed spread (often a fake constant) — is **not historically
knowable**, so an authored-friction historical broker is "horseshoe-throwing". Its
replacement is a **cost model**: an ordinary downstream **C9 graph of cost nodes**
that *approximates* the cost side a broker would produce, explicitly as an
approximation. The cost nodes live in `aura-std` (the cost-graph composite-builder
in `aura-composites`), never in the domain-free `aura-engine` (C14/C16). They are
**not** "additive on the R stream" in isolation: a cost node **reads the state it
depends on** — the price stream, a realized-volatility tap, a C11-recorded
interest-rate source, and the executor's per-cycle R-record / trade events — and
emits a **cost-in-R** stream that is subtracted from gross R to yield net R. Cost
attaches at the structurally correct grain: **per-trade** factors (commission, a
flat cost-per-trade) deduct from `realized_r` at close; **per-cycle-held** factors
(carry / funding / swap) accrue over the holding duration. The model **generalizes**
the existing scalar `round_trip_cost` / `net_expectancy_r` into a possibly
**state-dependent graph**: the scalar `round_trip_cost` is the degenerate
constant-per-trade special case, **subsumed** by the cost graph — the post-run
`summarize_r` fold no longer recomputes cost independently but folds the
cost-model's net-R stream into `net_expectancy_r` (one home for cost, no
double-count). Discipline: every cost factor is **either** a clearly-labelled
**stress-parameter** (e.g. a flat cost-per-trade is a breakeven-threshold probe)
**or** **data-grounded / falsifiable** (e.g. realized-volatility → slippage;
recorded interest-rate data via C11 → funding / swap). **Default simple;
complexity is earned per grounded factor.** Stacking unfalsifiable guesses
(over-modelling) is the anti-pattern.

**The research loop is pure feed-forward — compounding is removed.** Flat-1R-style
R accounting needs no equity, and with the Sizer gone there is **no equity → size
edge at all** in research: the loop is **feed-forward, maximally parallel and
deterministic (C1)**, the cost model a feed-forward subtraction on the R stream.
**Compounding is removed from research**: it is a **post-strategy money-management
transform** — a pure function of the per-trade **net-R sequence** and a
bet-fraction *f*, multiplicative and path / order-dependent, the **sole** source of
feedback — so compounding, Kelly-*f*, and drawdown-under-compounding are derived
**post-hoc, analytically, at the deploy / account layer** from the net-R
distribution. Consequently the `z⁻¹` fill-edge register **and** the
flat-1R-vs-compounding structural axis are **gone from the research loop** (with no
in-loop feedback there is nothing for a register to cut, so the C23-reordering
hazard the old text invoked to make the register mandatory no longer exists — this
is strictly C9 / C23-cleaner).

**Conviction-based risk allocation survives — as an R-aggregation axis, not a
Sizer.** Scaling risk by bias strength is **signal-side and R-denominated**.
Because per-trade R is **size-invariant**, conviction cannot be expressed by
scaling position size (that is invisible to R); it is expressed by **weighting the
per-trade R-contribution** in the R-equity: **flat** (sum of `realized_r`, sign
only) vs **conviction-weighted** (sum of `|bias| · realized_r`, sign + magnitude).
This is a **feed-forward, additive, order-independent research axis** (flat vs
conviction-weighted R-aggregation), distinct from the removed money-Sizer; it is
tested via the existing `conviction_at_entry` field and `conviction_terciles_r`
metric, and may sit in-graph or as a post-hoc fold.

**Money / real broker / cTrader Open API = a separate, later live / deploy-edge
concern — the only `belastbare` (reliable) ground truth, measured never modelled.**
Reliable friction statistics require **forward-trading against a real broker**
(e.g. cTrader Open API), and are non-stationary even then. This fits C11
(record-then-replay: real fills are recorded live, then replayable) and the
frozen-deploy invariant (C13: deploy = frozen bot + broker connection);
reconciliation with the real account is an **external I/O adapter** at the
recording / deploy edge, not an in-graph node. **This is the only place account
money appears.** (Currency-denominated *reference geometry* — pip value,
stop-distance-in-currency from the C15 `instrument_geometry` sidecar — is still
read at the **ingestion** edge to normalize a currency / pip cost factor into R;
notional size cancels in `cost_in_R = cost_in_currency / (size · stop_dist)`, so
the cost model is R-pure without ever holding *equity*. It is equity / account
money, not reference geometry, that lives only at the deploy edge.) The
broker-independent **position-event table** (`event_ts, action[buy/sell/close],
position_id, instrument_id, volume`) is the **deploy / reconciliation** record at
this edge (real volume), not a research artifact.

**Honesty principle.** The net-R curve under the cost model is a **research /
ranking tool and a hypothesis**; the **forward / live run against a real broker is
the ground truth**. The cost model *approximates*; it **never claims realism**.
`SimBroker` (the pip-equity, unsized-exposure node) is the **pre-reframe pip
ancestor** — with the net-R cost model it is **redundant as a quality measure** (R
supersedes pips). It is retained as a **legacy / simple optional pip yardstick**
(still wired in the `stage1-r` harness for an honest dual readout), **not part of
the new model and not to be expanded**.

**Forbids.** Putting **sizing or the stop in the strategy** (bias is unsized; the
stop-rule owns the stop, R is the unit); treating an **equity curve** (R or
currency) as the strategy's direct output; **putting a Sizer / currency size /
`volume` into the research loop** (size is a deploy concept; research is in R);
leaving a **dangling `size` input port** on the research executor (C8 wiring
totality forbids it — drop the port or drive it with a constant unit); **any
equity → size / equity → anything feedback in research** (the research loop is pure
feed-forward; compounding is a post-hoc money-management transform, not an in-loop
edge); modelling an **authored-friction "realistic broker" over historical data**
(real friction is not historically knowable — use the approximating cost model, and
treat the real broker as the live-edge ground truth only); **claiming the cost
model is realism** (it is an explicit approximation); **stacking unfalsifiable cost
guesses** (each cost factor is a labelled stress-parameter *or* data-grounded —
over-modelling is the anti-pattern); **computing cost in two homes** (the cost
graph owns cost; the post-run fold subsumes the old scalar `round_trip_cost`, never
double-counts it); expressing **conviction by scaling position size**
(size-invisible to R; conviction is an R-aggregation weight); making the
**position-event table the strategy's direct DAG output** (it is derived, not
emitted per `eval` — a decision instant may need >1 event, which C8 forbids) or a
**research** measure of signal quality (it is a deploy / reconciliation artifact);
**measuring signal quality in currency / account money** in the research loop at all
(account money lives only at the live deploy edge; reference geometry at ingestion
is not account money); baking a broker into the strategy or an **in-graph broker
subsystem** (the in-graph realistic broker is retired; the only broker is the
live-edge I/O adapter, C11 / C13); a signed-volume direction trick in the event
table (use `action`); storing `open_ts` (derive it from the opening event).

**Why.** A strategy's edge is a *procedure*, not a currency outcome ("focus on the
procedure, not the money"): the right primary question is *"how much R out per 1R
risked?"*, and R — defined by the stop — is the only account- and
instrument-agnostic, risk-normalized unit (pips are not). Separating **direction
(bias) from sizing from fill** is the decomposition every mature system converges
on (LEAN's Alpha → Portfolio-Construction → Execution is a near isomorphism;
backtrader, QSTrader, zipline all emit an unsized directional signal sized
downstream) — and aura pushes *sizing* and *fill* off the research loop for two
**distinct** reasons: **sizing** is off because per-trade R is **size-invariant**
(size carries no information in R — flat-1R is perfectly knowable, it just does not
matter), and **fill / friction** is off because real friction is **not
historically knowable** and therefore not honest over history. Keeping research
**pure feed-forward** leaves the signal-quality layer parallel and deterministic
(C1); the **only** real feedback (equity → bet-fraction) is **compounding**, a
closed-form, path-dependent transform of the net-R sequence, and therefore belongs
**after** the strategy, at the deploy / account layer, not as an in-loop register.
The **cost model as a C9 graph** keeps cost within the one Node / graph abstraction
(C9) and generalizes the scalar `net_expectancy_r` continuously, while the
**gross-R / net-R** split states the cost-drag honestly without inventing a unit.
Refusing the historical realistic broker is an **honesty** stance: the only
`belastbare` friction is **measured forward** against a real broker (cTrader Open
API) — the live deploy edge, the sole place account money and ground truth appear.
The DAG holds exactly one state at t and a node emits ≤1 record per `eval` (C8), so
the faithful per-cycle output is the **bias** (one value); position *events* are a
derived, deploy-side consequence. This supersedes the **realistic-broker /
currency / compounding / Stage-1-vs-Stage-2** framing of the #117 reframe (see the
2026-06-28 reframe note), which itself superseded the **exposure** framing of
cycle-0007 and the still-earlier "strategy's output is the position-event table"
framing.

**Reframe (2026-06-28, #116 — realistic broker retired; cost-model graph in R; money to the live edge).**
This is the live contract above. Ratified in an in-context design discussion
(reference issue #116). It **preserves** the durable spine — unsized bias stream
(sign = direction, magnitude = conviction), signal quality in R, the stop defining
R, and the decoupling of direction from sizing from fill — and the shipped Stage-1
realizations (the `Bias` node, `FixedStop` / `vol_stop`, `PositionManagement`,
`summarize_r` / `RMetrics` / `sqn_normalized`, SQN as ranking objective). The
**RiskExecutor composite survives in shape** (bias + price → stop-rule →
position-management) but with its **Sizer interior removed** and its `risk_budget`
argument dropped / vestigial (it sized the now-removed Sizer); the Veto remains an
optional documented seam. The contract **supersedes** the following, which were
**design intent, largely unbuilt** (#116 — the realistic broker and the whole
Stage-2 currency layer were never implemented; the Stage-1 R chain was) and are
retired:
- the **"realistic broker"** concept — an authored-friction historical broker —
  rejected as "horseshoe-throwing" (real friction is not historically knowable),
  replaced by the **cost model**: a composable C9 graph of cost nodes (in
  `aura-std` / `aura-composites`), approximating not claiming realism,
  generalizing / subsuming `round_trip_cost` into `net_expectancy_r`;
- the **"Currency P&L is Stage 2"** paragraph in full, the **Stage-1-vs-Stage-2
  hard sequencing gate**, and **currency / fixed-fractional / compounding** in
  research — compounding is now a post-hoc money-management transform of the
  net-R sequence at the deploy / account layer;
- the **`z⁻¹` register on the fill edge** and the **flat-1R-vs-compounding
  structural axis** as research mechanism — there is no equity → size edge in
  research, so no register and no such axis in the loop;
- the **Sizer in research** and currency **size / `volume`** — size is a deploy
  concept; the research executor is stop + position-management in R;
  `PositionManagement`'s `size` port is dropped or constant-driven (no dangling
  port, C8), its `size` field and the event table's `volume` are vestigial in
  research;
- the **"a broker is an ordinary in-graph node / no special external broker
  subsystem"** forbid, **for the live edge only**: in-graph brokers are retired
  outright, and the live broker is now an explicit **I/O adapter** at the C11
  recording / C13 deploy edge — not an in-graph node and not part of the research
  graph (the no-in-graph-broker-subsystem prohibition still holds inside the
  graph);
- the **position-event table as the realistic-broker input** and its
  first-difference-of-the-book (`deal = target − book − in_flight`) execution
  framing — the table (schema 0063 #114, derive 0068 #115) **survives** as the
  **deploy / reconciliation** artifact (real volume), not a research artifact and
  not "fed to a broker" in research.
Money, a real broker, and cTrader Open API are a **separate, later live /
deploy-edge** concern (C11 record-then-replay; C13 frozen-deploy invariant) — the
only `belastbare` ground truth, measured, never modelled. The honesty principle is
explicit: the net-R curve is a research / ranking hypothesis; the forward / live
run is ground truth. `SimBroker` is downgraded to a legacy / optional pip
yardstick, not to be expanded. (Terminology note: with Stage 2 gone, the
**"Stage-1"** label below is no longer one half of a two-stage gate — it survives
only as a historical cycle / identifier name, like the `exposure` → `bias` on-disk
alias, denoting the shipped feed-forward R chain.)

**Realization (cycle 0081 — cost-model graph, cycle 1, #148).** The first concrete
cost node + the net-R seam shipped. `ConstantCost` (`aura-std`) is an ordinary
downstream node (C9) emitting a 3-field cost-in-R record `{cost_in_r, cum_cost_in_r,
open_cost_in_r}` isomorphic to `PositionManagement`'s R-triple; R-pure
(`cost_per_trade / |entry − stop|`, notional cancels, no equity held). The scalar
`round_trip_cost` argument of `summarize_r` is **retired**: `summarize_r` folds a
co-temporal cost stream (positional 1:1 join — `cost[i]` is `record[i]`'s cycle) into
`net_expectancy_r`, one home for cost / no double-count, byte-identical to the old
cost = 0 baseline on an empty stream. The headline sink is **`net_r_equity`** =
`LinComb(4)[cum_realized_r, unrealized_r, −cum_cost_in_r, −open_cost_in_r]` → Recorder
(C8/C18), a sibling of `r_equity`, emitted only when a cost is authored. Wired on the
**run path** via `--cost-per-trade` (`stage1_r_graph`); sweep / walk-forward / mc pass
`None` (cost on the reduce-mode sweep path and cost in OOS pooling deferred — the
positional join holds only while one cost node fires in lockstep with PM). Deferred to
later milestone cycles: the general `CostNode` trait + multi-node cost-graph
composite-builder, data-grounded factors (realized-vol → slippage, recorded-rate →
swap), per-cycle-held accrual (carry / funding), and the conviction-weighting
R-aggregation axis. Decision log: #148.

**Realization (cycle 0082 — cost-graph composition, cycle 2, #148).** The cost graph
**composes**: a second, *state-dependent* cost node plus an aggregator prove that two
cost nodes sum into one net-R curve with `summarize_r` and the `net_r_equity` tap
structurally unchanged. `VolSlippageCost` (`aura-std`) charges `slip_vol_mult · vol /
|entry − stop|` in R, reading an **independent short-horizon realized-range** vol
(`RollingMax − RollingMin`, window distinct from the stop's own vol — scaling slippage
by the stop's vol would collapse cost-in-R to a constant, indistinguishable from
`ConstantCost`). `CostSum` (`aura-std`) is the cost-graph **output node**: it sums `N`
cost nodes' 3-field cost-in-R records **per-field** into one aggregate (`n = 1` is the
identity), so the seam consumes a single cost stream regardless of node count — one
home for cost, the positional join unchanged. **Co-temporality contract (load-bearing,
generalizes to all future factors):** since `summarize_r` positionally joins `cost[i]
↔ record[i]`, a cost node is gated **only by the PM trade-geometry**; any not-yet-warm
state input (the vol proxy warms later than PM) contributes **0 cost** that cycle
rather than withholding — the node still emits its row, so the cost stream stays
co-temporal 1:1 with the PM record. This makes co-temporality structural and
warm-up-independent, preserves the C18 golden, and is honest (no slippage estimate yet
→ no charge). `ConstantCost` satisfies it trivially; only state-dependent nodes need
the missing-factor → 0 rule. Wired on the **run path** via `--slip-vol-mult`,
composable with `--cost-per-trade` (their costs sum); sweep / walk-forward / mc pass
`None`. The 3-field cost triple is now restated by-convention across the two producers
+ `CostSum` (a compiler-unlinked lockstep) — to be unified by the still-deferred
general `CostNode` trait (now justified by two concrete nodes). Decision log: #148.

**Realization (cycle 0083 — CostNode trait + shared cost-record contract, cycle 3,
#148).** The deferred unifier ships. A new `aura-std/src/cost.rs` owns the cost-model
node abstraction: the 3-field cost triple is now **one source of truth**
(`COST_FIELD_NAMES` / `COST_WIDTH`, mirroring `position_management::{FIELD_NAMES,
WIDTH}`), read by both producers + `CostSum` + the CLI wiring — the cycle-0082
by-convention lockstep is **structurally gone** (four restatements collapsed to one).
The `CostNode` **factor trait** carries a cost node's only per-node difference — the
price-unit **cost numerator** (`cost_numerator(&mut self, &Ctx) -> f64`), plus
`extra_inputs` (default none) and `label`; everything else is the generic
`CostRunner<F: CostNode>` **adapter**, which holds the shared `cum`/`out` state and
implements `Node`, writing the co-temporality skeleton (geometry-only gating,
`numerator / latched` R-normalization, the closed/open charge, the running `cum`, the
3-field emit) **once**. `ConstantCost` and `VolSlippageCost` are now thin factors whose
`new()` returns `CostRunner<Self>`. Honours C9 (a `CostRunner<F>` is a plain downstream
`Node`, no runtime sub-object) and C23 (`name()` dropped as dead surface — `label()`
remains the non-load-bearing symbol). **Behaviour-preserving**: the builders emit
unchanged schemas, so the wiring / `net_r_equity` seam / `summarize_r` are untouched and
the `numerator / latched` token form is byte-identical; the existing suite passes
verbatim and two new CLI characterization goldens pin the exact flat/composed
`net_expectancy_r` (the prior tests only asserted `net < gross`). Decision log: #148.

**Realization (cycle 0084 — cost-graph composite-builder, cycle 4, #148).** Decision E
ships. A new `cost_graph(Vec<PrimitiveBuilder>) -> Composite` in `aura-composites` (the
C16 layer that couples the engine builder + `aura-std` nodes) is the cost-model graph's
**authoring primitive**: it fans the 4 PM-geometry inputs to `N` cost nodes, surfaces
each node's extra inputs (discovered via `schema().inputs[GEOMETRY_WIDTH..]`,
`GEOMETRY_WIDTH` now re-exported from `aura-std`) as `cost[k].<port>` composite roles,
sums them through `CostSum`, and exposes the 3-field aggregate. The CLI's manual
slot-indexed cost-wiring + the hardcoded `MAX_RUN_COST_NODES = 2` cap are deleted — the
composite handles arbitrary arity. **Behaviour-preserving** (C11): the composite inlines
at bootstrap to the same flat fan-in, so the cycle-0083 `net_expectancy_r` goldens are
byte-identical (four `aura-composites` unit tests pin the exposed role-set + output
triple, incl. arbitrary-arity per-node namespacing). Honours C9 (ordinary downstream
nodes), C16 (wiring stays out of `aura-engine`), C23 (role/port names are
non-load-bearing). **Carried debt (#152, for the deferred sweep-cost cycle):** the
`cost[k].<port>` index-namespacing is restated across `CostSum` / `cost_graph` / the CLI
(a build-validated lockstep, not the silent positional kind 0083 collapsed), and
`cost_graph` `.leak()`s runtime port names per build — fine for one-shot run-path
construction, but to be interned (the `COL_PORTS` production pattern, not the test-only
`.leak()`) before cost reaches the per-member sweep path. Decision log: #148.

**Realization (cycle 0085 — per-cycle-held accrual, cycle 5, #148).** C10's
"per-trade factors deduct at close; **per-cycle-held factors accrue over the hold**"
clause is first realized. A cost factor now declares *when* it charges via
`ChargeMode { AtClose, PerHeldCycle }` (a defaulted `CostNode::charge_mode()`, default
`AtClose`), read by the **one shared `CostRunner`** — not a second runner type (a
commission is intrinsically at-close, a carry intrinsically per-held-cycle; the timing
belongs to the factor, preserving the cycle-0083 "skeleton written once" win). The
`PerHeldCycle` arm accrues `per` into a per-position `acc` every held cycle, dumps the
accrued total into `cum` at close (resetting `acc`), and marks the open position via a
**growing `open_cost_in_r`**. The first accrual node, **`CarryCost`** (`aura-std`,
a `ConstantCost` twin differing only in `charge_mode()`), is a labelled stress
parameter — the flat base of the accrual family; a run-path `--carry-per-cycle` flag
pushes it into the existing `cost_graph`/`CostSum` aggregation (no new wiring; a
per-trade and a per-held-cycle node compose, the costs summing per-field). **Approach B
(honest bleed), achieved without a `summarize_r` fold change:** the headline
`net_r_equity` curve bleeds continuously over the hold because the bleed lives in
`open_cost_in_r`, which the `net_r_equity` tap already subtracts — so `summarize_r` and
the CLI `net_eq` wiring are **untouched**, and the cycle-0083/0084 `net_expectancy_r`
goldens stay byte-identical (the `AtClose` arm is the pre-cycle-5 eval tokens verbatim).
The 3-field cost record's semantics generalize cleanly: `cost_in_r` = cost realized this
cycle (into `cum`); `open_cost_in_r` = the open position's cost marked-to-market
as-of-now (would-be-close for `AtClose`, accrued-so-far for `PerHeldCycle`) — singly
counted, no `cum`/`open` double-count (the two are mutually exclusive per cycle). Honours
C9 (a `CostRunner<CarryCost>` is an ordinary downstream `Node`), C11 (byte-identity), C16
(factor in `aura-std`, `aura-engine` domain-free), C23 (`label()` carries the rate). The
B-vs-A discriminator (a growing intra-hold `open_cost_in_r`, invisible to the scalar
`net_expectancy_r`) is pinned by unit B-proofs + a `net_r_equity`-bleeds-over-the-hold
integration test. **Deferred (each its own #148 cycle):** a notional-based carry
(`price × rate`, reads the price tap), the calendar-aware **overnight swap** proper
(rollover-boundary timing, 3× Wednesday, long/short asymmetry — deploy-edge realism), and
sweep-path cost. Decision log: #148.

**Realization (cost-flag harness scoping, #153).** Cost flags are defined only
against an R-evaluator harness (the gross-R → net-R chain). On a non-R harness
(`--harness sma`/`macd`, which produce no R) a cost flag is a **usage error
(exit 2)**, not a silent no-op — refuse-don't-guess. Negative cost rates are
likewise rejected (exit 2) with a named diagnostic that identifies the offending
flag. CLI ergonomics only; the cost-model graph and the R math are untouched.

**Reframe (2026-06-23, #117 — exposure → bias, R as the signal-quality unit).**
[HISTORY — its R spine survives into the 2026-06-28 contract; its Stage-2 currency /
realistic-broker / register / flat-1R-vs-compounding portions are SUPERSEDED by that
reframe.] The contract was reframed from **exposure** (a signed fractional
position) to an **unsized bias** plus **R** as the signal-quality unit. The
realization notes below describe the **pre-reframe code** (`Exposure`, `SimBroker`,
pip-equity), retained as history and as the **ancestor** of the current chain:
`Exposure { scale }` is the ancestor of the unsized `bias` node, and `SimBroker`'s
pip integral is the ancestor of the R-evaluator (which additionally requires a
stop, since R is stop-defined). The `exposure → bias` rename, the RiskExecutor /
Sizer / Veto nodes, and the R-evaluator **landed in cycle 0065**; the
position-event schema (0063, #114) survives as the audit layer. Industry grounding
for this reframe: LEAN / nautilus_trader / backtrader / QSTrader / vectorbt /
zipline (see #117 decision log).

**Realization (cycle 0007).** [HISTORY — pre-reframe; `SimBroker` is now legacy per
the 2026-06-28 reframe.] The signal-quality half was realized at the substrate as
two `aura-std` nodes on the unchanged engine (the engine stays domain-free — it
routes only `f64` records). The **exposure stream** was realized as `Exposure {
scale }`: `clamp(signal / scale, -1, +1)`, one `f64` per fired cycle. The
**sim-optimal broker** was realized as `SimBroker { pip_size }`: a two-input node
(exposure, price) accumulating `prev_exposure · (price − prev_price) / pip_size`
and emitting cumulative pip equity — the exposure held *into* a cycle (decided at
t-1) earns that cycle's return (causal, C2); `pip_size` is held reference metadata
(C7/C15). An end-to-end harness (SMA-cross → `Exposure` → `SimBroker` → recording
sink) produced a recorded pip-equity curve, bit-identical across runs (C1).

**Realization (per-instrument pip channel, 2026-06, #22).** [HISTORY — pertains to
the legacy `SimBroker` pip channel.] `SimBroker`'s `pip_size` is sourced **per
instrument** from the recorded geometry sidecar (`instrument_geometry`, over
data-server's `symbol_meta`), at the ingestion / source edge (never `Aura.toml`,
never `Ctx`); the engine stays domain-free. (The original cycle-0022 form was a
Rust-authored vetted floor `InstrumentSpec { pip_size }` + `instrument_spec(symbol)`;
cycle 0074 removed that floor — see the C15 note — once the sidecar geometry made it
redundant for the real path. Refuse-don't-guess on absent geometry.) The honesty
rule is **refuse, don't guess**: a real-data run for a symbol with no vetted spec
is a usage error (`exit 2`). Threaded through the CLI `aura run --real` path; the
manifest broker label records the looked-up pip.

**Realization (position-event schema, cycle 0063, #114).** [Survives as the
**deploy / reconciliation** schema per the 2026-06-28 reframe — no longer a
broker-input research artifact.] A closed `PositionAction { Buy, Sell, Close }`
enum + the `PositionEvent` row (`event_ts`, `action`, `position_id`,
`instrument_id`, unsigned `volume`; no `open_ts`; direction *is* the action) live
beside `RunMetrics` as a post-run value type (not a per-`eval` node — C8) — both in
the `aura-analysis` crate since cycle 0079 (#136); originally `aura-engine`, see the
C16 cycle-0079 note. `action` serde-encodes as a bare `i64` (Buy=0, Sell=1,
Close=2), the C7 scalar column form, with an out-of-range code rejected on read.
The table stays broker-independent.

**Realization (cycle 0065 — Stage-1 R signal quality, #119/#126/#127/#128/#129).**
[The R spine here is the live model; the *Stage-2 deferral* clause at its end is
SUPERSEDED by the 2026-06-28 reframe — there is no Stage-2 currency / compounding
layer; cost is now the cost-model graph and money lives only at the live deploy
edge. "Stage-1" reads as a historical identifier, not a gate half.] `Exposure →
Bias` renamed the unsized strategy output (node + output field); the persisted
`exposure_sign_flips` metric key (serde alias), the `SimBroker` `exposure` input
slot, and the on-disk `exposure` **tap** label retain the old name. The
strategy-output **param namespace** (the `Bias` instance, its `bias.scale` knob,
the `bias_scale` manifest param) was completed in #134. A **stop-rule** defines 1R:
`FixedStop` (a triggered-constant primitive) and a `vol_stop(length, k)`
**composition** `k·√EMA(Δ²)` (a composition of `Mul` / `Sqrt` primitives).
**`PositionManagement`** (`aura-std`) is the stateful heart: it latches the
entry-cycle stop distance as the immutable R-denominator, marks against the
one-cycle-lagged fill (no look-ahead, C2), and emits a **dense 14-column per-cycle
R-record** (one row per eval, C8; the trade ledger is the `closed_this_cycle`
subset, the R-equity is `cum_realized_r + unrealized_r`). The **`Sizer`** (`size =
risk_budget / stop_distance`, flat-1R) was the feed-forward sizing seam, and **R is
size-invariant** — scaling `risk_budget` leaves every `realized_r` unchanged
(pinned by a RED test); per the 2026-06-28 reframe the Sizer and `size` / `volume`
are removed from research (size is a deploy concept). **`summarize_r`** is a
post-run fold (sibling of `summarize`, **not** an in-graph node) → `RMetrics` (E[R],
SQN, win-rate, profit-factor, max-R-drawdown, conviction terciles, net-of-cost);
per the 2026-06-28 reframe it folds the cost-model's net-R rather than recomputing
a scalar cost. The **RiskExecutor** ships as a public `aura-composites`
composite-builder (`risk_executor(StopRule, risk_budget)`) with a
`StopRule{Fixed,Vol}` **structural axis** (C20); per the 2026-06-28 reframe its
Sizer interior and `risk_budget` arg are dropped / vestigial, the **Veto** stays a
documented seam, not a runtime node. The layer is operable from the CLI: `aura run
--harness <sma|macd|stage1-r>` — a compile-time selector over Rust-authored
harnesses (C9/C17) — folds `summarize_r` into `RunMetrics.r`, the `stage1-r`
harness fanning one bias into both `SimBroker` (legacy pip) and the RiskExecutor
(R); an `r_equity` tap charts the by-trade R-equity. Composites live in the
dedicated `aura-composites` crate, so `aura-engine`'s runtime dependency stays
`aura-core`-only and `aura-std` is an `aura-engine` `[dev-dependencies]` (the graph
stays acyclic).

**Realization (cycle 0066).** SQN is the operational single-number objective for
ranking a Stage-1 sweep family by signal quality — C12 **axis-2 (argmax-metric)**
over the C18 family store. `metric_cmp` (`aura-registry`) learns the
higher-is-better R metrics `sqn`, `expectancy_r`, `net_expectancy_r`; a member
with no `r` block sorts last (`NEG_INFINITY`). `aura sweep --strategy stage1-r`
produces the rankable R family, each member folding `summarize_r` into
`RunMetrics.r`. The default grid varies **only the signal** (`fast` / `slow` SMA
lengths), holding the stop and sizing fixed: `risk_budget` is R-invariant and
`bias.scale` is sign-only under flat-1R, and — load-bearing — the **stop defines
1R**, so varying it would change what R *means* per member and break cross-member
SQN comparability. Each swept member's manifest records the fixed R-defining params
beside the floated knobs (reproducible from its own manifest, C18).

**Realization (cycle 0067, #130 + #135).** **#130 (SQN100):** `RMetrics` gains
`sqn_normalized = (mean_R/stdev_R)·√(min(n, 100))` — the n-normalized "SQN score"
(`SQN_CAP = 100`), turnover-robust where raw `sqn` rewards trade count. It is an
**opt-in** rank key (`Metric::SqnNormalized`); raw `sqn` and the default ranker
stay byte-unchanged, and the field carries `#[serde(default)]` (C18). Below the cap
(`n ≤ 100`) it equals raw `sqn` exactly. **#135 (stage1-r `--trace`):**
`stage1_r_sweep_family` persists each member's equity / exposure / r_equity under
`runs/traces/<n>/<member_key>/` via the same `persist_traces_r` the single run uses;
per-member `--trace` is symmetric across all three sweep strategies.

**Realization (cycle 0068, #115 — position-event derive).** [The derivation
survives as the **deploy / reconciliation** layer per the 2026-06-28 reframe — its
"realistic brokers consuming it" goal is retired; money is the live-edge concern.]
`derive_position_events(record, instrument_id) -> Vec<PositionEvent>` (in
`aura-analysis` since cycle 0079, #136; originally `aura-engine`, beside
`summarize_r`) is the **first difference of the executed book**: a pure post-run
reduction over the `PositionManagement` dense record (read positionally as
type-erased `Scalar`s, C7 SoA — no in-graph node, so the hot path stays
domain-free, C14), emitting a `Buy` / `Sell` at each open and a `Close` at each
exit, a reversal (or stop-then-same-cycle reopen) emitting **Close then the
opposite open at one `event_ts`** (close first — the C8 ">1 event per instant" case
that forces a *derived* table, not a per-`eval` output). The close sizes the
**actual book** (the closed position's stored volume), never an exposure delta.
`instrument_id` is a caller-supplied scalar (`aura-analysis` depends only on
`aura-core`). A position open at window end emits its open with **no synthetic
`Close`** (the table records actual executed events; `summarize_r`'s force-close is
for the R metric only). The `r_col` ⟷ PM-record lockstep is guard-pinned for
`direction` too.

### C11 — Generalized sources; record-then-replay determinism boundary
**Guarantee.** A source is anything that produces timestamped scalar streams —
market data (`data-server`) and non-financial sources (e.g. a news-agent node
emitting a bias) are treated identically. Anything nondeterministic, external,
or slow (LLM/news/web) is materialized into a recorded, timestamped stream
*before* it enters the engine; backtest replays the recording, live computes
fresh in real time and records it for future backtests. A bias enters as a value
held until the next event (firing policy A).
**Forbids.** Any live external call *inside* a backtest replay.
**Why.** It is the only model compatible with reproducible backtests — LLM calls
are nondeterministic and far too slow per-cycle. Per `~/.claude/CLAUDE.md`,
external LLM (IONOS) calls happen only at the recording/live-source edge, with
explicit per-session consent, never inside a sim.

### C12 — The atomic sim unit and the four orchestration axes
**Guarantee.** The atomic unit is `(frozen topology + param-set + data-window +
RNG-seed) → deterministic run → metrics`. Parameters are typed, ranged, runtime
values injected at graph build (no recompile per param-set; the optimizer sees a
generic vector of typed ranges). Raw data is shared read-only across sims via
`Arc<[T]>` (data-server is built for this). Four axes orchestrate the atomic
unit: (1) param-sweep (grid/random), (2) optimization (argmax metric),
(3) walk-forward (rolling in-sample optimize + out-of-sample test),
(4) Monte-Carlo (N seeded realizations perturbing input). **MC = sweep over
seeds**; each realization is itself deterministic given its seed.
**Realization (cycle 0041).** The eager ingestion of cycle 0011 is no longer the
only path. `Harness::run` is re-typed to a **producer seam** — a `Source` trait
(`peek`/`next`, object-safe) the k-way merge drives — and a streaming
**`M1FieldSource`** (`aura-ingest`) pulls a data-server window lazily, borrowing
**one** `Arc<[M1Parsed]>` chunk per pull (zero-copy *within* a source) and
decoding each `Scalar` on demand: the **source ring** is resident O(one chunk),
not O(window length) — the measured `resident_records()` predicate, a *per-source*
bound, not whole-process RSS. (Data-server's `FileCache` retains each window's
parsed chunks read-only for the pass — ~56 B/record — so process residency is
O(records-touched); that is the **replay-many** sharing C12 wants — one window
parsed once across a sweep family — not a leak. The single-pass cost is tracked as
#95.) The eager `load_m1_window`/`close_stream` path is kept for bounded loads (the
gap closes by a streaming path *existing*, not by deleting the eager one). **Realized
(2026-06-29, family cycles shipped).** cross-*sim* `Arc<[T]>` sharing — one window shared
zero-copy across many disjoint sweep sims — is now in force: the sweep / Monte-Carlo /
walk-forward families (axes 2–4 above) build their members over **one** shared
`Arc<DataServer>` (one `FileCache`), so a window is parsed once and every member's
`M1FieldSource` borrows the same cached `Arc<[M1Parsed]>` chunks
(`crates/aura-ingest/src/lib.rs:316`). The single-pass *parse* cost stays tracked as #95.
**Realization (cycles 0028, 0049).** Axis 1 (param-sweep) is built. `GridSpace`
(0028) enumerates a cartesian lattice over discrete per-slot value-lists;
`RandomSpace` (0049) draws `N` seeded points over typed continuous `ParamRange`s
(I64 inclusive `[lo,hi]`, F64 half-open `[lo,hi)`), validated against the
param-space before any run. Both implement the `Space` trait the disjoint
`sweep` / `run_indexed` core is generic over, so either enumeration runs through
one execution path (C1: results in enumeration order, not completion order). The
seeded sampler reuses the bit-stable `SplitMix64` as a **code-path-disjoint**
instance from the data-edge seed RNG (the source-seam firewall, #52/#71: they
share only the `u64` type, never a path).
**Forbids.** Baking a specific search strategy (Bayesian/genetic) into the
primitive — those are pluggable policies atop the atomic unit; recompiling on a
param change.
**Why.** A stable primitive + orchestration axes keeps "wahnsinnig schnell"
(embarrassingly parallel across the unit) cleanly separated from search policy.
Seed-as-input reconciles Monte-Carlo with C1. The "frozen topology" of the
atomic unit is one harness instance, selected by the harness's **structural
axes** (C20); the structural experiment matrix is the outer orchestration over
this dimension, the tuning sweep the inner (C19/C20).

### C13 — Hot-reload is authoring-only; deploy is frozen
**Guarantee.** A node/strategy is authored as a native Rust `cdylib`,
hot-reloaded during the authoring loop (Rust-ABI; host and node built with the
same toolchain). The live/deploy bot is a statically-linked, versioned, frozen
artifact.
**Forbids.** Hot-swapping a running live bot; loading third-party / foreign-
toolchain plugins.
**Why.** Hot-reload makes the research loop fast; a live artifact must be frozen
and reproducible (audit trail: this bot = this commit). A sweep pays no
hot-reload tax — params are runtime data (C12), so the cdylib loads once.

### C14 — Headless core, two faces
**Guarantee.** The engine is a UI-agnostic library. Two faces sit on it: a
**programmatic/CLI** face (the primary surface for the LLM and automation —
author a node, run a sim/sweep, emit structured metrics) and a **visual** face
for human exploration. Visualization is only a downstream consumer node on the
streams.
**Forbids.** Any UI/pixel knowledge inside the engine.
**Why.** The LLM drives programmatically, the human visually; a headless core
serves both. The visual face is the **playground** (C22) — a **web frontend
served from disk-persisted traces** (revised 2026-06, issue #101: the engine
writes recorded traces to disk, a browser charts them; supersedes the earlier
egui-native / in-process-zero-copy-from-SoA direction). It is staged after the
runnable substrate but is core to aura's identity, not optional. The pivot keeps
this contract's core intact — and arguably tightens it: a browser reading a
serialized trace file is a stricter "no UI knowledge in the engine" than egui
reaching zero-copy into the live SoA columns.

### C15 — Resampling-as-node; sessions/calendars
**Guarantee.** A resampler is a node (finer stream → coarser bar stream),
clock-sensitive, emitting a completed bar only at the boundary (C2). Calendars
and instrument specs are metadata (non-scalar, beside the hot path); session
*context* is exposed as scalar streams via a `SessionNode` (`bars_since_open:
i64`, `in_session: bool`, `session_open_ts: timestamp`). "3rd 15m candle after
session open" is then a plain node checking `bars_since_open == 3`.
**Forbids.** Streaming the calendar; special-casing session logic outside the
stream model.
**Why.** Keeps the line consistent — everything a signal needs arrives as a
stream; reference data feeds source/session nodes from beside the hot path.
**Realization (instrument specs, 2026-06, #22 → #124).** The "instrument specs are
metadata" half of this contract is realized by the **recorded geometry sidecar**:
`instrument_geometry(server, symbol)`, over data-server's `symbol_meta`, returns
neutral broker-agnostic `InstrumentGeometry` (the raw provider JSON never enters this
repo) — non-scalar reference data held beside the hot path, keyed by symbol, feeding
the sim-optimal broker's pip divisor (C10). The real-path pip is sourced from
`InstrumentGeometry.pip_size`; refuse-don't-guess on absent geometry. **History:**
cycles 0022/0063 first carried a Rust-authored vetted floor (`InstrumentSpec` — a
single `pip_size`, later a six-field deploy-grade row `instrument_id`,
`contract_size`, `pip_value_per_lot`, `min_lot`, `lot_step`, `quote_currency`, plus
`tick_size`/`digits`), cross-checked against the sidecar geometry. **Cycle 0074
removed that floor**: with the sidecar geometry supplying the real-path pip, the
authored table was redundant, so `InstrumentSpec` / `instrument_spec` / the vetted
list were deleted. The speculative deploy-grade fields and the override/floor tiers
of the #124 hierarchy (tier 1 authored override, tier 3 authored floor) are
**deferred** until a consumer — the cost model's currency→R normalization (C10), or
the live deploy edge (real broker) — needs them, read from the sidecar geometry then. The `quote_currency` runtime-type transition is likewise
deferred to that consumer; #124 collapses to **recorded sidecar geometry → refuse**
in the meantime.

**Realization (SessionNode — `bars_since_open` only, by design; #154).** The
session-context half of this contract ships **one** scalar stream, not the three the
Guarantee lists: `Session` (`crates/aura-std/src/session.rs`) emits only
`bars_since_open: i64` (tz-aware, DST-correct, baked Frankfurt open). This is a deliberate
narrowing pinned in the node's own contract — *"there is no separate in-session bool gate,
`bars_since_open` alone is the contract"*: a downstream `EqConst(== N)` gate subsumes the
`in_session: bool` stream (pre-open instants give `<= 0`, which never match), and
`session_open_ts: timestamp` has no consumer. The two streams the Guarantee names remain
the original design intent, **deferred until a consumer needs them** (default-simple;
forward-queued as #154). The stream model is untouched — session context is still exposed
as scalar streams fed from beside the hot path.

### C16 — Engine / project separation; three-tier node reuse
**Guarantee.** aura is the reusable **engine**; each research project is a
separate external repo that depends on aura via cargo (the game-engine / game
split). Node reuse is cargo-native, in three tiers: **`aura-std`** (universal
blocks, ship with the engine) / **shared node crates** (cross-project-reusable,
their own repos, pulled as cargo git deps) / **project-local `nodes/`**
(experimental, project-specific). A reusable node is an `rlib` dependency; the
hot-reload unit stays the project-side `cdylib` that composes it (consistent
with C13). Concretely a project is a **Rust crate** — a cdylib library of node /
strategy / experiment blueprints — plus a static `Aura.toml` (project context:
data paths, instrument/pip metadata, default broker & window, runs dir). During
research the `aura` host loads and runs it (C13 hot-reload); for deploy the
chosen strategy + broker freeze into a standalone binary. **A project is
therefore always a Rust program built on the engine.** Beyond the node-reuse
tiers, the engine workspace also carries non-node crates — `aura-engine` (the
run loop), `aura-cli` (the `aura` binary), and **`aura-ingest`** (the
data-source ingestion edge, cycle 0011, where the `data-server` external tree
enters). **Dependency policy (amended 2026-06-10, cycle 0029).** Dependencies
are admitted by deliberate, **per-case review** — what a crate pulls in weighed
against what it buys — with **particular scrutiny for anything that enters the
frozen deploy artifact** (C13: this bot = this commit). Well-established
standard crates (`serde`, `rayon`, …) pass that review and are used wherever
they do the job, including in the bot. There is **no blanket zero-dependency
commitment** and **no blanket admission**; hand-rolling what a vetted standard
crate already does is the anti-pattern, not the dependency. This strikes the
original *"zero-external-dependency by commitment"* clause and the
*"`aura-ingest` is the sole external-dependency firewall"* framing; `aura-ingest`
remains the data-source ingestion edge, no longer a dependency wall.
**Forbids.** Project-specific signals in the aura repo (it keeps at most
example/fixture nodes under `examples/` for its own tests); a multi-project
manager inside aura; a bespoke node registry/marketplace (cargo + Gitea *is* the
package mechanism). (Dependency admission is governed by the per-case policy
above, not a blanket ban.)
**Why.** The engine/game split keeps the engine sharp and reusable while each
project versions its own research with its own forward-queue. Promotion
(local → shared → std) is the ordinary Rust reuse gradient, no new mechanism.
**Realization (cycle 0079, #136 — the analysis leaf becomes its own crate).** The
trading-domain **analysis** layer — the post-run reductions that are pure functions of a
run's recorded data (`RunMetrics`, `RMetrics`, the R reduction `summarize_r` /
`r_metrics_from_rs`, the position-event table `PositionEvent` / `PositionAction` /
`derive_position_events`, and the multiple-comparison hurdle math `inv_norm_cdf` /
`expected_max_of_normals`) — is extracted out of `aura-engine`'s `report` module into a
dedicated **`aura-analysis`** crate (deps: `aura-core` + `serde` only; `serde_json` is a
dev-dependency, test-only). This sharpens the engine↔domain seam: the run loop's crate no
longer *defines* the trading metrics — it `pub use`-re-exports them for source
compatibility this cycle, and still hosts the trace-coupled `summarize`, `RunReport`,
`RunManifest`, and the columnar trace utils. The non-node engine-workspace crates are now
`aura-engine` (run loop), `aura-cli` (`aura` binary), `aura-ingest` (ingestion edge),
`aura-composites` (composite-builder convenience layer), and `aura-analysis` (post-run
domain reductions + selection provenance). Behaviour-preserving (C1): every serde shape
is byte-identical (C18 goldens green).
**Realization (cycle 0080, #136 — engine-side extraction complete).** The last
trading-domain types still *defined* in `aura-engine`'s `report` module — the
selection-provenance types `FamilySelection` / `SelectionMode` — also move to
`aura-analysis` (the only non-cyclic home: `RunManifest.selection` needs the type and the
engine already depends on `aura-analysis`; the registry/CLI reach them unchanged via the
re-export). Behaviour-preserving (C1; suite 665/0). **Settled (2026-06-27, user-ratified
in the #136 thread):** the `aura-registry` `Metric` / `metric_cmp` / deflation vocabulary
**stays in the registry by design** — the registry is the trading *selection* layer, not
the domain-free engine, and its deflation statistics branch irreducibly on metric
identity (R-bootstrap vs. `total_pips` dispersion floor); forcing genericity would be a
one-implementor abstraction. **Deferred to the World / C21 layer (explicitly *not* #136):**
making `RunReport` generic over a metric type and turning the orchestration/registry layer
into a reusable domain-agnostic substrate — until then `RunReport` / `RunManifest` /
`summarize` stay trace-coupled in the engine.

### C17 — Authoring surface
**Guarantee.** All *logic* — nodes, strategies, **and experiments/harnesses** —
is authored in native Rust through **Claude Code + the skills pipeline**: the
human describes, Claude writes the Rust, builds it, runs it via the `aura` CLI,
and reports metrics. Declarative config (`Aura.toml`) carries only **static
project context** (data paths, instrument/pip metadata, defaults, runs dir),
never logic. aura ships **no embedded coding-LLM**. IONOS LLMs are used only as a
*runtime data source* (news-agent bias, C11), gated by per-session consent,
never in the code path.
**Forbids.** An in-app LLM chat that generates node code inside aura; using
IONOS (weaker models) as the authoring brain.
**Why.** LLMs author Rust well in Claude Code — that is the fix to RustAst's
failure; making weaker models the coding brain reintroduces the very problem.
Keeps aura's scope an engine + playground, not an LLM-IDE.
**Refinement (2026-06-29, #109 resolved — node *logic* is Rust; *topology* is data,
C24).** "All logic … including experiments/harnesses … is Rust" governs
**computation** — a node's math, a meta-program's control flow — and that stays
native Rust (the RustAst lesson, sharpened: a small LLM cannot reliably *apply* a
computational DSL, and a strong one does not need one). It does **not** bind
**topology** — which node feeds which, the structural axes, the param-space — to
Rust *source*. Topology is a **serializable data value** the World owns (C24): a
non-Turing static DAG over the closed, compiled-in node vocabulary (C8/C16),
carrying no computation. This is *not* the RustAst trap re-opened: that trap is a
DSL the author must *apply*; a topology-data blueprint is machine-generated / owned
and applied by no one. The no-DSL guard therefore stands unbroken — it forbids a
*computational* experiment / strategy language, not a static topology-data format.
The experiment-matrix *generator* may stay Rust (C20); its *output* is
topology-data.

### C18 — Project management: one repo = one project, plus a run registry
**Guarantee.** Management has two planes. (1) **Code & forward-queue:** git
(commit = identity; the frozen bot *is* a commit) + Gitea (ideas/hypotheses as
the forward-queue, a research thrust = a milestone, the
`idea → experimental → validated → deployed` label gradient). (2) **Experiments
& results:** an Aura-native **run registry** — one record per run = a *manifest*
(node-commit + params + data-window + seed + broker profile) + *metrics*,
queryable, with *lineage* (composite ← signals; run ← inputs). Determinism
(C1/C12) makes a run reproducible from its tiny manifest, so the registry stores
manifests + metrics and re-derives full results on demand. Depth: **structured**
(promotion/status, lineage, run-diff).
**Forbids.** Storing results not reproducible from a recorded manifest;
duplicating git/Gitea inside aura; a multi-project workspace manager.
**Why.** Comparing experiments over time is the heart of the research loop and
has no home in git/Gitea; determinism makes a structured registry cheap.
Sequencing: the walking skeleton emits a manifest + metrics per run from day
one; the registry/index is a later milestone over manifests that already exist.

**Realization (cycle 0029 — the flat run registry).** The experiments-&-results
plane shipped as `aura-registry`: an append-only JSONL store (`runs/runs.jsonl`),
one serde_json `RunReport` (`RunManifest{commit, params, window, seed, broker}` +
`RunMetrics`) per line, with a typed read-path (`load`) and best-first ranking
(`rank_by`/`optimize`). C9 holds — the registry depends on `aura-engine`, never
the reverse.

**Realization (cycle 0045 — lineage as related records, #70).** The *lineage*
depth is now realized as a **family store**: a sweep / Monte-Carlo / walk-forward
run (the C12 axes) is persisted as a *set of related records* — each a
`FamilyRunRecord` (a `RunReport` stamped with its `family` + `run` + `kind` +
`ordinal`) — in a sibling JSONL (`families.jsonl`), leaving the flat `runs.jsonl`
path and its `append`/`load`/`rank_by`/`optimize` API byte-for-byte unchanged.
the user-facing `family_id = "{family}-{run}"` handle is **derived** from the
stored `family` name plus a per-name `run` index (assigned as a numeric max+1 —
not a content hash; re-running the same family mints a fresh id). `group_families` is the round-trip that re-derives a family
from the stored links (re-listable / rankable as a unit — C21). The manifest *is*
the re-derivation recipe (#71): no input-stream blob / path / payload enters a
record, and a member's window is **producer-supplied** via
`Source::bounds()`/`window_of` (eager or streamed → byte-identical lineage),
never a materialized-`Vec` scan at the call site. CLI surface: `aura mc`,
`aura runs families`, `aura runs family <id> [rank <metric>]`; `aura sweep` /
`walkforward` / `mc` persist via `append_family` with an optional `--name`.
**Realization (cycle 0078 — cross-instrument family + instrument lineage, #146).**
The comparison axis (C12) is realized as a `FamilyKind::CrossInstrument` family:
`aura generalize` runs one candidate across an instrument list and persists the M
per-instrument runs via `append_family`, each member self-identifying through a new
first-class `RunManifest.instrument` lineage field (serde-widened with
`skip_serializing_if`, so legacy lines and every non-cross-instrument path stay
byte-identical — C14/C23). The cross-instrument *generalization score* (worst-case R
floor + sign-agreement + per-instrument breakdown) is a **recomputable aggregate**
over those members, not a persisted family-level record — distinct from #144/#145's
per-winner selection annotation on `RunManifest.selection`.
Deferred (Non-goals): content-addressed identity + replay-dedup; the "run-diff"
depth and ranking families against each other (cross-family, vs. within-family);
and a live producer for the flat `runs.jsonl` standalone-run path — no CLI command
writes it (sweep/walkforward persist to the family store; `aura run` does not
persist). **Resolved (#73, 2026-06): retired** — `aura runs list` / `rank` dropped;
families (C21) subsume standalone over-time comparison. The `aura-registry` flat
lib API is retained: `rank_by`/`optimize` keep live consumers (`optimize` backs
walk-forward's in-sample step, `rank_by` backs `runs family … rank`); `append`/
`load` (the flat-store half) remain public API with **no in-tree caller** after
this retire — tested, available to external consumers, a latent dead-code surface
a later sweep may revisit. **Unknown-id contract (ratified, Runway fieldtest
2026-06).** `aura runs family <id>` treats an unknown-but-well-formed id as an
*empty family* (prints nothing, exit 0) — the same treat-as-empty discipline as
`Registry::load` reading a missing store as `Ok(empty)`, and deliberately
distinct from the retired `list`/`rank` exit-2, which is argv-shape rejection
*before* any store access, not a found-nothing lookup. Tightening to a non-zero
`no such family <id>` exit (typo-safety) is an available future UX choice, not a
current contract.

**Refinement (2026-06-29 — a generated run's topology lives in / is addressed by
its manifest, C24).** The manifest identifies a run's topology today only via
`commit` (this engine + a hand-coded harness) — sufficient while harnesses are a
finite hand-coded menu. Once the World **generates or structurally searches**
topologies (C21/C24), `commit` no longer identifies the graph, so the manifest must
**carry or content-address the topology-data** to stay the re-derivation recipe
(C18's "reproducible from a recorded manifest"). This pulls the previously-deferred
**content-addressed identity** non-goal forward as the natural home for a generated
topology's identity. The format and the carrier are C24's design; recorded here as
the reproduction requirement it must meet.

### C19 — Bootstrap: blueprint → instance (recursive)
**Guarantee.** Construction is a distinct phase, recursive at every level. Each
node type has a **factory** `params → sized concrete node` (e.g. `SMA(length)`
sizes its ring buffer). A **blueprint** is the param-generic, input-role-generic
graph-as-data produced by running a Rust builder (C9); it carries *free* numeric
params (declared ranges) and *free* input roles. The **bootstrap** binds
`(blueprint + param-set + data bindings + seed)` into a concrete, **frozen
instance** — buffers sized, topology fixed. This is precisely the "wiring /
graph build" that C7 ("sized at wiring", "topology frozen per sim") and C12
("params injected at graph build") already reference. The same machinery applies
recursively up to the harness (C20). A sweep builds many instances from one
blueprint; instances are disjoint (C1).
**Forbids.** Params that change topology (a topology change is a *different*
blueprint — Fork A, C7 "frozen"); resizing buffers after bootstrap; running a
sim against an un-bootstrapped blueprint.
**Why.** Separating the param-generic blueprint from the param-bound instance is
what makes one strategy reusable across a whole sweep and lets the optimizer
mutate "the 20" by *rebuilding* an instance (cheap; no recompile, C12) instead
of rewriting code. Naming the build phase makes the implicit "wiring" of C7/C12
explicit.
**Refinement (Construction-layer milestone, 2026-06-05).** This binding is a
**compilation**: the param-generic, named blueprint (source) is lowered to a flat,
type-erased **flat graph** (C7) **wired by raw index, not by name**
(`Edge { from, to, slot, from_field }`): composite boundaries dissolve entirely,
and field / role names are demoted to **non-load-bearing** debug symbols (as
`FieldSpec.name` already is). "No recompile" above means no **Rust /
cdylib** rebuild (C12/C13: the cdylib loads once); re-deriving an instance per
param-set is a cheap **graph re-compilation**, not a code recompile. Naming the
build phase a compilation makes its successor explicit: the flat graph is the target
of behaviour-preserving optimisation (C23).
**Realization (cycle 0016 — param-set injection).** The bootstrap now binds an
injected param-set, realizing C12's "params injected at graph build (the optimizer
sees a generic vector of typed ranges)" and C19's "factory `params → sized node`"
*literally*: a blueprint leaf is **value-empty** — `BlueprintNode::Leaf` holds a
`LeafFactory { name, params, build }` recipe, not a built node — and the value lives
only in the injected vector (no baked default), so the blueprint stays a pure
param-generic recipe. (Renamed in cycle 0024: `BlueprintNode::Primitive` holds a
`PrimitiveBuilder { name, schema, build }` — the recipe now carries the full
signature, see the C8 0024 realization; `bootstrap_with_params`/`compile_with_params`
moved onto `Composite` when `struct Blueprint` collapsed into it, see the C19 0024
realization.) `bootstrap_with_params(Vec<Scalar>)` /
`compile_with_params` **build each leaf through its own constructor** (the single
sizing/validation gate) from its kind-checked slice **while lowering** (build-then-
wire), consuming the vector slot-by-slot in the *same* depth-first walk
`param_space()` projects — so the two share one traversal (subsuming the #34 dual-
traversal hazard) and the value reaches the node at the slot the sweep enumerates.
Arity is checked up front (`param_space().len()`); a wrong-kind or wrong-length
vector is a typed `CompileError::{ParamKindMismatch, ParamArity}` (the typed-value
check C8 deferred). The lowering/edge/source rewrite is structurally unchanged, so
the **flat graph stays bit-identical** for a given point (C23, the correctness
invariant). The value *domain* (e.g. `length ≥ 1`) stays the constructor's own
`assert`; the search-range is still the run's (#32/C20). One value-empty leaf
detail for C22: the blueprint view (pre-run, param-generic) labels a leaf by **bare
type** (`PrimitiveBuilder::label`, was `LeafFactory::label`, → `[SMA]`) — the
value-bearing `SMA(2)` of C8's render-
label refinement now appears only in the *compiled* view (built nodes, `Node::label`).
**Realization (cycle 0017 — blueprint render = main graph + definitions).** The
`aura graph` blueprint view (C9 graph-as-data, #13) renders the **authored
structure** as a *program with subroutines*: a flat **main graph** wiring the
harness with each composite shown as a **single opaque node** `[name]`, plus a
`where:` section that defines each **distinct** composite type **once** (its
interior with named input-entry nodes and outputs folded onto their producers as
`name := …` bindings (render refined through cycles 0019–0022; originally
`[in:k]`/`[out]` port markers); deduped by `name()`, collected
recursively so nested composites are opaque nodes with their own definitions). This
supersedes #13's original cluster-box model and is the durable split this view
realizes: **blueprint = source** (composites as named subroutines, body once) vs.
**compiled = inlined machine form** (C23, boundary dissolved). The substantive cause
for retiring cluster boxes is a real renderer defect — `ascii-dag` 0.9.1's subgraph
level-centering rounds sibling x-positions with `/2`, overlapping wide sibling
labels (width/parity-sensitive, no config/padding dodge surviving unequal-width
siblings); the **flat** layout is collision-free, and both views now build flat
graphs only. The model also scales (blueprint size tracks top-level wiring, not
inlined node count) and removes #13's nested-composite `unimplemented!` (the
definitions pass recurses). The interactive *enter/focus* counterpart (a composite
collapsed to a navigable node) is the playground's, parked as a separate concern
(#37); the static CLI keeps the all-at-once definitions form (#38).

**Realization (cycle 0024 — the root is the fully-bound composite; the flat graph is
a named type).** `struct Blueprint` is **deleted**: the root graph IS a `Composite`,
and `compile_with_params`/`bootstrap_with_params`/`param_space` are its methods.
What distinguished the root — its bound data sources — is now a property of its input
roles: `Role` carries `source: Option<ScalarKind>` (`None` = an open interior port,
wired by the enclosing graph; `Some(kind)` = a bound ingestion feed). A composite is
**runnable iff every root role is bound** (C3: sources bind at ingestion only); an
open root role is a compile-time `CompileError::UnboundRootRole`. So the "main graph"
is no longer a separate kind — only the composite all of whose roles are
source-bound, governed by the same conditions as any other node. Compilation now
targets a **named type**: `compile` validates structurally **pre-build** (via
`signature()`, no node constructed — an ill-typed wiring is caught before any build
closure fires) and emits `FlatGraph { nodes, signatures, sources, edges }` — the
C23 flat graph, now first-class — which `Harness::bootstrap` consumes (kinds/firing
from the carried signatures, buffer depth from `lookbacks()`). The per-flat-node
signature travels beside the node, so `bootstrap` reads it without a built-node
`schema()` call (which no longer exists, see the C8 0024 realization).

**Realization (cycle 0026 — graph render redesign: model + WASM-Graphviz viewer,
#51).** `aura graph` no longer renders ASCII. The render path is now two pieces:
a read-only **model serializer** (`aura_engine::model_to_json`, iteration 1) that
walks the root composite + every distinct composite type into a deterministic,
hand-rolled JSON model (C14, golden-tested; the engine's last hand-rolled JSON
writer after `RunReport::to_json` moved to serde in cycle 0033; the
swapped-param mis-wire property moved here from the old compiled-view test), and a
**self-contained HTML viewer** (`aura-cli::render::render_html`, iteration 2) that
inlines that model, the ported prototype viewer JS, and a vendored Graphviz-WASM
blob into one page emitted to stdout. Layout/SVG happen in the browser via
WebAssembly — aura ships no layout engine and stays a serializer (C9: graph-as-data,
no `eval`/build on the path). The viewer is a render asset (C10 — no node/strategy
logic, no DSL); it labels every input pin from the model's now-real names (C23
debug symbols, named in cycle 0027) and colours wires by the four scalar base
types (C4). This **retires `ascii-dag`** and its adapter (`graph.rs`), the
`--compiled`/`--macd` flag plumbing, and the invented `#Sf`/`:=`/`histogram →`
notation — superseding the cycle-0017 flat-ascii model and its renderer-defect
workaround. The DOT/SVG are Graphviz-version-dependent and not golden-tested; the
deterministic JSON model is the asserted contract.

**Realization (cycle 0034 — structural-constant bind: a knob *removed* from
param_space, #55).** `PrimitiveBuilder::bind(slot, value)` adds the **third** param
category beside the topology factory-arg (C7/C19) and the tuning param (the
cycle-0016 value-pin): a **structural constant**. The cycle-0016 binding *pins* a
value in the injected vector while the knob **stays** in `param_space` (a tuning
param the sweep varies); `bind` instead **removes** the slot from `param_space`
entirely — the knob is gone, not fixed. The discriminator is the #55
deform-vs-tune test: a value whose variation yields another valid point of the
*same* strategy is a **tuning param** (stays in `param_space`); a value whose
variation *deforms* the strategy into a different one (e.g. the `2` of an
"SMA2-entry" bound to its two-candle construction) is a **structural constant**
(bound out), so a sweep never enumerates deformed strategies as valid family
members. Mechanically `bind` shrinks the builder's declared param surface
(`schema.params`) and wraps its build closure to re-splice the constant at its
original positional slot; the construction layer
(`collect_params`/`lower_items`/`param_space`/`compile_with_params`) is
**byte-unchanged** — both dock sites already key off `builder.params()`, so the
shrink propagates for free, and chained binds reconstruct the correct positional
vector because each layer computes its slot index relative to the param list it
sees. C23 is unaffected: `bind` resolves the param **name** to a position at
**authoring** time (the by-name authoring address space, the 0032 amendment) and
the flat graph stays wired by raw index — the name never reaches it. The
complementary question — exporting a **named frozen** strategy (all/most knobs
bound) as a reusable blueprint *value* — is deferred (#60); `bind` ships only the
per-knob overlay, no registry (C9/C10 intact).

### C20 — Strategy ↔ harness; the harness is the root sim graph
**Guarantee.** A **strategy** is a reusable composite-node blueprint (C9):
broker-, data-, and viz-independent, with inputs declared as named **roles**
(symbol-agnostic where possible) and the **bias** stream (C10) as output.
A **harness** (the experimental setup) is the **root sim graph** — sources bound
to the strategy's input roles + the strategy + attached broker node(s) + sinks —
and is itself produced by the bootstrap (C19). A harness *instance* is C1's
disjoint unit (RustAst's "root scope"). The harness has **two kinds of
parameterization**: **structural axes** (which strategy, which instrument(s),
which broker(s), which window) whose variation selects *different* instances —
the **experiment matrix**; and **tuning params** (the strategy's numeric params)
swept *within* a fixed structure (Fork A). The same strategy blueprint is reused
across backtest, sweep, visual workspaces, and the frozen live bot — each a
different harness. **Both strategy and harness/experiment are authored in Rust**
via builder APIs (C17); the experiment matrix is ordinary Rust control flow
(loops/conditionals), not a config schema. Ontologically, a node (incl. a
strategy composite) is an **open** fragment — free input roles + ≤1 output
(C8/C9) — that does not run alone; a **harness is the closed root graph**: a
strategy with its input roles bound to sources and its output terminated in
broker/sink nodes, under a clock. A harness is therefore **not a node** (no free
inputs, no output; it does not fit `eval`) — it is the *closure that runs*, C1's
disjoint unit / the root scope. Harnesses do not nest as nodes; the World (C21)
orchestrates them as objects.
**Forbids.** Embedding data sources / brokers / sinks inside a strategy; a
declarative experiment mini-DSL (logic is Rust — C17); modelling the harness as
a node (it is the closed root scope, not an open composable node).
**Why.** Reusability needs the strategy to be a context-free blueprint that many
harnesses embed. Modelling the harness as a root graph keeps it within the one
Node/graph abstraction (C9) and makes "10 strategies in one environment" and
"one strategy × N instruments" plain nested loops over the structural axes. Rust
authoring (not config) preserves full programmatic power — conditional/adaptive
matrices, generated axes, custom wiring — and avoids re-introducing the DSL trap
C17 rejects.
**Realization (GER40 session-breakout blueprint milestone, 2026-06-17 — refs
#94/#96/#97, spec 0051).** Made concrete on the first real-data strategy: a
**hand-wired `FlatGraph` is not a shippable strategy** — it carries no
`param_space()`, so the World families (sweep / walk_forward / compare, C21)
cannot consume it without a hand re-author (the friction the GER40 deep-dive
fieldtest surfaced). The **canonical shippable form is the `Composite`
blueprint** (the authoring/source level, C9/C19); the `FlatGraph` is only its
compiled substrate (C23). The breakout now ships as
`ger40_breakout_blueprint(bar_period, …)` whose `param_space()` is exactly its
tuning knobs (`{entry_bar.target, exit_bar.target}`); the **bar period is a
construction argument** binding Resample + Session together — a structural
matrix axis (C12: a different period is a *different strategy*, not a sweep
point), never a `param_space` entry, so a sweep cannot desync the two clocks.
This is the concrete instance of the structural-axis-vs-tuning-param split
above (and `delay.lag`, a C8 structural constant, is bound out of the space).
**Refinement (2026-06-29 — experiment-matrix *members* are topology-data, C24).**
"The experiment matrix is ordinary Rust control flow, not a config schema" holds for
the **generator** — the loop / conditional that *enumerates* structural variants may
stay Rust (C17). But each **member** it yields is a **topology-data value** (C24),
not a hand-coded blueprint nor engine-baked source: a structural axis selects among
*data* blueprints the World constructs, mutates, and serializes. This is what makes
structural variation **first-class and searchable** rather than a hand-enumerated
menu — `HarnessKind` and the per-strategy `*_sweep_family` functions in `aura-cli`
are the pre-C24 scaffolding (topology-as-engine-source, a C16 tension), retired as
C24 + the project-as-crate layer land.

### C21 — The World: the meta-level is the product
**Guarantee.** Above the harness sits the **World** — the project's program /
"game" (C16: a Rust crate). Within it, **harnesses are dynamically constructible,
first-class objects**: meta-programs (walk-forward, sweep, optimize,
Monte-Carlo — C12's axes) construct harness instances at runtime via the
bootstrap (C19), run them disjointly in parallel (C1), aggregate / compare their
results, and discard the transient instances. A walk-forward rolls windows →
bootstraps a harness per window → stitches out-of-sample equity + parameter
stability into one meta-result. Orchestrating *families* of harnesses is
**first-class**, not a headless afterthought; the run registry (C18) is the
World's memory.
**Forbids.** Relegating multi-harness orchestration (walk-forward / sweep /
comparison) to second-class headless-only status; treating the single backtest
as the product.
**Why.** What happens *within* one harness — backtest a strategy → equity — is
commodity; every quant system covers it. aura exists for the meta-level: a
programmable space where harnesses are dynamically built and families of them
orchestrated and explored. The deterministic single-harness engine (C1–C20) is
the **substrate**; the World is the **product**.
**Refinement (2026-06-29, #109 resolved — the World owns topology as data, C24).**
"Harnesses are dynamically constructible, first-class objects" is sharpened: their
**topology is a serializable data value the World owns** (C24), not Rust source.
This is the precondition for the World's full ambition — it can **generate**,
**mutate**, **structurally search** (genetic / NEAT-style graph operators, the open
search-policy thread), **compare**, and **serialize-for-reproduction** (C18)
*families that vary structure*, not only numeric params. A param sweep over a fixed
skeleton is the degenerate case; varying the skeleton itself needs
topology-as-value. The game-engine principle (C16) made literal: the engine owns the
scene / blueprint graph as content, instantiates and runs it, the way aura owns
harness topology as data.

### C22 — The playground is a trace explorer; sinks are the recording mechanism
**Guarantee.** The World is a *program*: nothing is displayable until it runs,
and its harnesses are transient machinery (built, run, discarded — C21). The only
durable, displayable substance is the **recorded trace**, captured by **sinks** —
pure consumer nodes (C8) that persist a stream (equity, position events, a node's
output) into the run registry (C18). **Displayable = exactly what a sink
recorded**; with no sink, only input params + summary metrics remain. The
**playground** is therefore an **execution viewer / trace explorer**, and it
plays *any* harness (it is the harness *player*, not a harness, never bound to a
default one): before a run it shows the program *structure* (graph-as-data, C9) +
param knobs; during a run, live sink streams; after a run, recorded traces +
metrics from the registry — including meta-views (stitched walk-forward, sweep
surfaces, multi-strategy / instrument comparison). Observability is **explicit
and selective** — you instrument what you want to see; the choice of sinks is
part of the experiment. The engine ships sample harnesses *with* sinks so a
newcomer sees a populated trace immediately; a fresh project is empty until
built, run, and instrumented.
**Forbids.** A scene-editor model that assumes a persistent populated world;
constructing or wiring topology in the UI (topology is Rust + hot-reload,
C9/C17 — the UI reflects the live graph-as-data and tunes runtime params via
sliders, C12); retaining un-sinked data; binding the playground to a single /
default harness.
**Why.** A program has no scene-at-rest to inspect; what persists is what it
records. Making sinks the one recording-and-observability mechanism keeps "what
can I see?" answerable by "what did I instrument?", and keeps the engine
UI-agnostic (C14). Live param tuning (runtime values, no topology change —
C12 / C19 Fork A) gives the interactive feel without a wiring DSL.
**Realization (cycle 0006).** Sinks-as-recording-mechanism is realized at the
substrate level: a recorded trace is exactly what a recording node pushed out of
the graph (no engine recording registry; the constructing World holds each
recording node's destination). The engine's single `observe: usize` affordance is
removed — `Harness::run` returns `()` and recording is a node-side concern, so one
run records *many* streams (one per recording node) instead of exactly one row.
Recorded streams are sparse and timestamped (a record per fired cycle, tagged
`ctx.now()`), matching a trace of timestamped events (C18). No new contract; the
`Harness` API change (observe removed, `run -> ()`) is recorded here.
**Realization (the web-from-disk visual face, #101).** The C14/C22 web seam shipped
(amendment 3b56efb): a recording run persists each drained tap as a columnar (SoA,
C7) `ColumnarTrace` to `runs/traces/<name>/` (`aura-registry::TraceStore`, beside the
run registry's `runs.jsonl`), reachable via `aura run [--real <SYM> …]
--trace <name>`; `aura chart <name> [--panels]` reads them back, aligns all taps on a
synthetic-union timestamp spine via the post-run `join_on_ts` (C3 — not in-graph;
C1 — pure, no live external call), and emits a static self-contained uPlot page
(vendored like `render_html`'s Graphviz-WASM). The engine stays headless (C14):
encoding lives in `aura-engine` (`ColumnarTrace`, struct→JSON only), file I/O in
`aura-registry`, rendering in `aura-cli` (`render_chart_html` + `chart-viewer.js`).
First cut only — overlay (per-series y-scale) / timestamp-aligned panels; the served
page injected the chart *series* but no run-context header (the run manifest is
persisted on disk in `index.json`, deliberately not surfaced in that first cut — a
ratified scope call). The header (#102) and serve-time decimation (#108) landed in
cut 2 (see the served-page-hardening amendment below); the families-comparison view
landed too (#107). A local server and replay-clock controls remain open (see "Open
architectural threads").
**Amendment (family-member traces, #104, d3cb5f8).** The disk-trace layout extends
to family runs: `aura sweep|mc|walkforward --trace <name>` persists *each member* as
a nested standalone run-dir `runs/traces/<name>/<member_key>/`, reusing
`persist_traces`/`TraceStore` verbatim (engine untouched — persistence is a
side-effect inside each per-member closure). The `member_key` is content-derived and
deterministic — sweep: the *varying* axes rendered as a filesystem-portable
directory component (`<axis-name>-<value>` tokens joined by `_`, charset
`[A-Za-z0-9._-]`, case-less values, length-capped with an FNV fallback), MC
`seed{N}`, walk-forward `oos{ns}` — never
a runtime ordinal, because members run in parallel under the engine's `Fn + Sync`
HOFs, where a counter would be schedule-dependent (C1); concurrent `TraceStore::write`
targets disjoint member dirs, so it is lock-free. Opt-in: without `--trace`,
stdout/registry are byte-unchanged. Because `TraceStore` resolves `<name>/<member_key>`
as a subpath, `aura chart <name>/<member_key>` charts any single member with no
view-side change — the write-side precondition for the family-comparison **view**
(overlay / small-multiples across members) — now realized (#107; see the
families-comparison amendment below).

**Amendment (real-data family source, #106, 8e5d14b).** The family runs gain an
**opt-in real-data source axis**: `aura sweep|walkforward --real <SYMBOL> [--from
<ms>] [--to <ms>]` streams real M1 close bars from the data-server archive (the #71
`M1FieldSource` seam — C6 record-then-replay: a pre-recorded archive, never a live
call mid-sim) instead of the built-in synthetic stream, per family member. A
CLI-side `DataSource` provider (synthetic | real) supplies each member's source,
pip (resolved from the recorded geometry sidecar (`instrument_geometry`) — C10
refuse-don't-guess; a symbol with no recorded geometry exits 2 before any data
access), window, and — for walk-forward — its `WindowRoller` sizes
(synthetic: bar-index 24/12/12; real: fixed calendar-time 90d/30d/30d in ns; CLI
flags for these are deferred). The engine, ingest, and registry are untouched (C9 —
the whole change is in `aura-cli`); the synthetic path is byte-unchanged.
**MC is excluded** (#106 Fork A): its seed varies a *synthetic* price-walk
realization (C12 axis 4), which is undefined over real data's single realization —
`aura mc --real` refuses with exit 2; a real MC needs a bootstrap-resampling axis
(its own cycle). The real walk-forward member key `oos{from.0}` widens from a small
synthetic bar index to an epoch-ns integer (still portable, collision-free). The
built-in demo strategy's **length grid is likewise data-kind-dependent**
(`DataSource::strategy_lengths`): synthetic keeps the short lengths that fit the
18/60-bar demo streams (trend SMA `{2,3}×{4,5}`, MACD `2/4/3`), real uses realistic
M1 lengths (trend `{50,100}×{200,400}`, MACD `12/26/9`) so the cross is a genuine
trend signal over tens of thousands of bars, not noise. This — like the roller
sizes — is a *demo-strategy calibration* patch; the real answer is project-authored
strategies (C9: a project crate owns its own grid), deferred to the project-env
work.

**Amendment (families-comparison view, #107, 4c64feb).** The family-comparison
**view** the #104 amendment left open is now realized. `aura chart <name>`
classifies the name on disk (`TraceStore::name_kind`: a top-level `index.json` is a
single run; else member subdirs with `index.json` are a family; else not-found) and,
for a family, overlays **one tap** (default `equity`, `--tap` to pick) of **every
member** in one self-contained uPlot page: `build_comparison_chart_data` emits one
labelled series per member, all on a **single shared y-scale** — members measure one
identical quantity, so a shared scale is what makes them comparable, unlike the
single-run overlay of *different* taps (per-series scales). The series align on the
same union-ts spine via `join_on_ts`, so **one mechanism yields three correct
readings**: sweep/MC members share the data window (a true overlay), walk-forward
members are disjoint OOS windows (null-complementary → the stitched curve). The view
consumes a **set of runs** (`&[FamilyMember]` from `TraceStore::read_family`, sorted
by key for determinism, C1); a family is its *first producer* — the deliberate first
step toward the programmable analysis meta-level (C21) **without** rebuilding the
orchestration axes (that is its own later cycle). Name resolution is made a **total
function** by the write-guard `TraceStore::ensure_name_free`, called once per tracing
command (`run`/`run --harness macd`/`run --real` → Run; `sweep`/`mc`/`walkforward` → Family)
to refuse cross-kind name reuse — so the one ambiguous on-disk state (a name used by
both a run and a family) is unreachable. Every error path exits 2, never panics
(C18/C10 refuse-don't-guess). Engine untouched (C9/C14): the whole change is
`aura-registry` (the resolver + classifier + guard) + `aura-cli` (the builder +
`emit_chart` name-kind branch + `--tap`); the single-run page is byte-unchanged when
no `--tap` is given. **Still deferred:** multi-column tap selection
(`build_comparison_chart_data` projects column 0 — the comparison taps in scope are
single-column f64; #47); and the orchestration-composability rebuild (sweep/MC/WFO as
composable meta-level tools).

**Amendment (served-page hardening — cut 2, #102 + #108, 476342d).** The two deferred
follow-ons the #107 amendment named are realized, both confined to `aura-cli` (engine
+ registry untouched, C14). **Run-context header (#102):** a `ChartMeta`
(kind/name/commit/window/broker/seed/taps, + member count for a family, + bound params
for a single run) is built from the `RunManifest` in
`build_chart_data`/`build_comparison_chart_data`, injected into
`window.AURA_TRACES.meta`, and rendered by a pure `buildHeader` in `chart-viewer.js`
(headless `.mjs`-guarded, like `buildCharts`). The family window is the **span** across
members `(min from, max to)` — the only reading that labels a disjoint walk-forward
family's true OOS coverage (it collapses to the shared window for sweep/MC).
**Decimation (#108):** a pure `decimate(ChartData, buckets) -> ChartData` min-max
transform thins the served page to ≤ ~2·`CHART_DECIMATE_BUCKETS` spine slots
(per-bucket min+max, nulls preserved), applied in `emit_chart` on both paths — a
multi-year M1 family now renders a few-thousand-point page instead of 100s of MB; full
recorded data stays on disk (view-only), and the walk-forward null-fill page collapses
for free. Deterministic (C1): pure bucketing over the sorted-deduped spine, no-op under
budget. **Refinement (#111, 2957561):** the per-bucket reduction is **tap-aware** — an
*envelope* series (equity) keeps min/max, but a *bounded level* series (the C10
bias level — pre-reframe: exposure — ∈[-1,+1]) reduces by per-bucket **mean** (a `ReduceKind` on each `Series`,
set by `reduce_for_tap`). Without it min/max made every multi-year exposure a solid
-1..+1 band (each bucket straddles many sign flips), reading as per-point oscillation;
the mean shows the net/duty-cycle level. Deferred refinements: rendering the min/max
envelope honestly as range bars / OHLC rather than a polyline (#112); a `--width`
budget flag and true intra-bucket min/max ordering for the non-default continuous
x-mode (#110).
**Refinement (2026-06-29 — the visual face is orthogonal and optional; it does not
motivate C24).** The blueprint data format (C24) is required by C21 (generation /
structural search) and C18 (reproduction of a generated run) **at the CLI level,
autonomously** — independent of any visual surface. The playground's form (web or
egui, **editor or viewer, or not built at all**) is therefore **downstream and
optional**, and topology authoring does **not** live in it. This severs an earlier
conflation that treated the playground as where topology-editing would live. C22's
"never a scene editor" is refined to its true content — **no persistent
scene-at-rest** (a World is a *program*, C21) — which does **not** forbid a future
blueprint *editor* over the World-owned topology-data (C24); should the playground
ever become one, the data format is already the substrate it round-trips, not a new
mechanism.

### C23 — Graph compilation and behaviour-preserving optimisation
**Guarantee.** The bootstrap (C19) is a **compilation**: it lowers a param-generic,
named **blueprint** (the authoring source — nodes, composites, strategy, harness;
C8/C9/C20) into a flat, type-erased **`FlatGraph`** — the frozen runnable
instance (C7) — **wired by raw index, not by name** (`Edge { from, to, slot,
from_field }`). Composites are **inlined** at this step (C9): the composite
*boundary* dissolves entirely (there is no composite in the flat graph). The
blueprint's field / input-role *names* are **non-load-bearing** — the wiring
resolves by index, and names survive at most as **informative debug symbols**
(exactly as `FieldSpec.name` already is, C8), kept for tracing / rendering (C9
graph-as-data, #13) but carrying no run semantics. The flat graph is then the target of **behaviour-preserving
optimisation** — any transform that leaves every observable sink trace
bit-identical (C1 is the correctness invariant) — on two levels:
- **intra-graph** (within one graph): **common-subexpression elimination** —
  two identical nodes (same type, same bound params, same input source) merge to
  one with output fan-out, so fractal composition (C9) costs no redundant compute —
  and **dead-node elimination** — a node on no path to a sink is dropped.
  Pure-consumer sinks (C8, no output) are never CSE candidates, so their
  out-of-graph side effects are preserved by construction.
- **across the sweep family** (over the family of flat graphs one blueprint yields
  under a param sweep, C12): the sub-graph whose nodes carry **no swept param** and
  whose inputs are **all sweep-invariant** (transitively from the sources — same
  data window) is **loop-invariant** w.r.t. the sweep and is computed **once**; its
  output is materialised as a recorded stream (C11) shared read-only across the
  sweep instances (`Arc<[T]>`, C12). Only the parameter-dependent suffix re-runs
  per sweep point — loop-invariant code motion over the sweep loop.
**Forbids.** Any "optimisation" that changes an observable trace (it breaks C1);
optimising over a representation that is not graph-as-data — an opaque
trait-object interior (the rejected nested-composite reading of C9) cannot be
analysed or rewritten across its boundary, so it forecloses both levels.
**Why.** Determinism (C1) is not merely an audit property; it is the **licence**
for these rewrites — a behaviour-preserving transform is only meaningful because
"same input → bit-identical run" makes "same result" decidable, and a
sweep-invariant sub-graph is reproducible enough to compute once and share. The
flat graph representation (C19) is the necessary condition: only a flat, inspectable
graph-as-data — with each node's declared param-ranges (C8) — lets the compiler
identify identical sub-expressions, dead nodes, and the sweep-invariant frontier.
This is precisely why a composite is an inlining (C9) and not a runtime
sub-engine: the flat graph is what makes the whole graph optimisable.
**Status (Construction-layer milestone).** The flat graph *representation* —
blueprint → inline / compile → flat instance — is the load-bearing substrate built
**first**; the optimisation passes (intra-graph CSE/DCE, sweep-invariant
hoisting) are **deferred**, behaviour-preserving follow-on work, each gated by a
"flat graph with pass ≡ flat graph without pass, bit-identical run" test (C1). The
sweep-level pass further presupposes per-node param declarations (C8 — landed in
cycle 0015 as `name`+`kind`; the swept *range* still pending #32/C20) and the sweep
orchestration itself (C12/C21).
**Amendment (cycle 0032 — param-namespace injectivity is part of the compilation).**
The bootstrap-as-compilation now includes one structural gate:
`check_param_namespace_injective` over the `param_space()` name projection (see C9 /
C12-C19), run before name resolution. It reads the **boundary** name projection — the
authoring address space — and rejects a duplicated path (`DuplicateParamPath`). Node
names stay **non-load-bearing**: they qualify the param path at construction and are
dropped at lowering (the flat graph is wired by raw index, unchanged). The check makes
the by-name *authoring address space* injective; it does **not** make names
load-bearing in the flat graph.

### C24 — The blueprint is a serializable, World-owned data value (the topology data format)
**Guarantee.** A **blueprint** — the param-generic, named graph-as-data a Rust
builder produces (C9/C19) — is a **first-class serializable data value** with a
stable format, and the engine has a **load path** (data → blueprint → `FlatGraph`,
C23), not only the Rust-builder construction path. Topology is therefore a **value
the World owns** (C21): constructed, **generated**, mutated, **structurally
searched**, serialized for **reproduction** (C18), and (optionally) rendered or
edited by a visual face (C22). `model_to_json` (C9) is the **existing reverse half**
(blueprint → data, for render); this contract closes the loop. The format carries
**topology + structural axes + the param-space**, referencing nodes by their
**compiled-in type identity** — the closed, typed node vocabulary of `aura-std` + the
project `cdylib` (C8/C16) — and carries **no node logic** (computation is compiled
Rust, C17). It is **non-Turing-complete by construction** (a static DAG, no control
flow), so it structurally cannot become a second RustAst.
**Forbids.** Putting **node logic / computation** in the format (the RustAst trap —
logic is Rust, C17); a **Turing-complete or control-flow-bearing** blueprint format
(it is static data; the *generator* that emits it may be Rust, C20, but the artifact
is not a language); a **by-name node marketplace / registry** (the vocabulary is the
compiled-in closed set referenced by type identity — the format is not a distribution
mechanism, C9/C16); treating the **visual face (C22) as the motivation or the
authoring brain** (the format is required by C21/C18 at the CLI level, independent of
any UI); a **generated / searched run whose topology is not recoverable from its
manifest** (C18 reproduction); baking topology as **engine source** compiled into the
binary (the hard-wired `aura-cli` harnesses are pre-C24 scaffolding, a C16 tension,
retired as this lands).
**Why.** The World (C21) is the product, and it cannot orchestrate, **structurally
search**, or **reproduce** *families that vary topology* while topology stays opaque
Rust source baked into the engine. The **game-engine principle** (C16's engine / game
split) makes it concrete: the engine is compiled native systems; the "game" is
**content** — and topology *is* content, a data value the engine owns, serializes,
instantiates, and mutates, exactly as a game engine owns its scene / prefab /
blueprint graphs. The #109 fork was never "*who types the wiring*" — a small LLM
cannot reliably *apply* a DSL and a strong one does not need one, so a
manifest-as-authoring-surface dies on both horns (C17). It is "**is topology a
compile-time *source* artifact or a runtime, serializable, World-owned *value*?**",
and determinism (C1) + reproduction (C18) + structural search (the open search-policy
thread) force the **value**. The engine already treats topology as a runtime value
(C9 graph-as-data, C19 "cheap graph re-compilation, not a code recompile", C23 the
flat graph as the optimisation target); C24 only adds that the value **serializes out
of Rust and loads back in**.
**Status (2026-06-29; first cut shipped — cycle 0087 / #155, `d5602ec`;
construction service shipped — cycle 0088 / #157).** The
**principle** is settled (this contract, ratified in an in-context design discussion
— the #109 resolution). The **first cut now ships**: a `Composite` blueprint
serializes to a **canonical, versioned** data value (`format_version` envelope,
omit-defaults JSON) and **loads back** (data → blueprint → `FlatGraph`) via
`blueprint_to_json` / `blueprint_from_json` (`aura-engine::blueprint_serde`),
referencing nodes by **compiled-in type identity** through an **injected resolver**
whose concrete closed `match` over the `aura-std` vocabulary lives outside the engine
(`aura-std::std_vocabulary`) — the engine stays domain-free, no node registry
(invariant 9). Acceptance met: a serialized blueprint runs **bit-identical** (C1) to
its Rust-built twin. `model_to_json` (C9) remains the render half; this closes the
loop for the round-trippable vocabulary. **Cycle 0088 (#157) adds the
introspectable construction service**: a declarative, replayable by-identifier
op-script (`aura graph build` / `introspect` over a JSON op-list, the engine
`GraphSession` / `replay`) builds a runnable blueprint through validated ops — the
engine's construction gates split into *eager* per-op checks (name resolution, edge
kind-match, the double-wire arm, param bind, and acyclicity — a `connect` that would close a
cycle is rejected eagerly, #161) and *holistic* finalize checks (wiring
totality, param-namespace injectivity, root-role boundness), **both cadences calling
the same extracted predicates** (`edge_kind_check`, the shared resolution helpers,
`check_root_roles_bound` — no second validator) — plus build-free introspection over
the closed vocabulary (types, a node's ports/kinds, a partial document's unwired
slots). It **emits** the #155 blueprint; acceptance met: a graph built purely through
the ops compiles identical (C1) to its Rust-built twin, an invalid op is rejected at
the op naming the cause, introspection answers without a build. The engine `Op` stays
serde-free (the wire DTO is CLI-side); the vocabulary's enumerable companion
(`std_vocabulary_types`) lives in `aura-std` (no registry, invariant 9). **Cycle
0089 follow-up (#161 / #162, after the cycle-0088 fieldtest).** The eager acyclicity
gate above (`GraphSession::closes_cycle`, a reachability check at the closing
`connect`) closed a fieldtest-found hole where `graph build` accepted a non-DAG
op-list (invariant 5 / C9). **Lockstep:** it is a *second* home of invariant-5
alongside the bootstrap Kahn sort (`harness.rs`, `BootstrapError::Cycle`); the two
must co-evolve when the explicit delay/register node — invariant-5's sole legal
feedback — lands, or a valid delay-feedback graph the bootstrap accepts would be
rejected at construction. The holistic *finalize* faults now also read
**by-identifier** (#162): `finish()` translates the index-carrying `CompileError`s
(`UnconnectedPort` / `RoleKindMismatch` / `UnboundRootRole`) into by-identifier
`OpError` variants — still *calling* the unchanged holistic gates (the
no-second-validator lockstep preserved), only translating their result.
**Cycle 0090 (#156) codifies the forward-compat two-tier discipline.** Tier-1
(additive-optional) is serde-default silent-ignore with no `format_version` bump
(a new optional field defaults to prior behaviour, C1) — now proven by
`unknown_optional_field_is_tolerated_byte_identically` (an unknown optional key
loads byte-identically and runs bit-identically). Tier-2 (must-understand: a new
node type, edge semantics, or structural-axis kind) bumps `format_version` so an
old reader refuses cleanly (`LoadError::UnsupportedVersion` / `UnknownNodeType`,
already green). The per-section required-flag scheme is deferred (no current
Tier-2 section to validate it; recorded on #156).
**Milestone delivered — 2026-06-30, cycle 0090 — the serializable format + loader + construction service.** A green end-to-end milestone fieldtest (`fieldtests/milestone-topology-as-data/`) proves the full author → serialize → load → construct → introspect → reproduce story from the public surface alone, 0 behavioural bugs; the round-trippable format (#155), the construction service (#157), and the forward-compat two-tier discipline (#156) all shipped. Polish filed forward: op-script grammar docs + a stale example (#163), the canonical trailing-newline / `Composite` value ergonomics (#164), CLI discoverability of `build`/`introspect` (#159).
**Deferred to the World / project-as-crate cycle** (moved out of this milestone — they address a problem that only arises once the World *generates* runs from blueprint-data, which does not exist yet: `commit` still fully identifies every hard-wired topology): content-addressed topology identity in the manifest (#158, C18) — premature until a run is built *from* a serializable blueprint; retiring the pre-C24
hard-wired `aura-cli` harnesses (`HarnessKind`, `run_stage1_r`, `*_sweep_family`) once
the project-as-crate layer lands (#159, paired with #157's data-authoring surface). **Out of the first cut's round-trippable set**
(deliberate; fails clean as `UnknownNodeType`, never a silent wrong graph): recording
sinks (capture an `mpsc::Sender` — runtime identity, not param-generic data, C19) and
construction-arg builders (`LinComb` / `CostSum` / `SimBroker` / `Session` —
structural-axis args, a C20 concern), additively addable later (#156). Still
sequencing-coupled to the **project-as-crate authoring layer** (`aura new` /
`Aura.toml` / `cdylib` loading, C16/C17) and the **composable-orchestration** thread
(#109): topology-as-data is the substrate both stand on.

---

## Open architectural threads not yet resolved

- **Playground & World UI surface** — the playground is core (C22); the surface
  is a **web frontend served from disk-persisted traces** (revised 2026-06, issue
  #101; not egui). Settled for the first cut: raw per-tap trace persistence to
  disk (columnar/SoA form, C7) + serve-time `join_on_ts` alignment + a static
  self-contained HTML chart page (uPlot, vendored like the `render_html`
  Graphviz-WASM blob), feeds overlaid or timestamp-aligned in panels. The
  families-comparison **view** — `aura chart <family>` overlaying one tap across a
  family's members on a shared y-scale — shipped (#107, see the C22 amendment).
  Still open: richer comparison meta-views (sweep-surface heatmaps, cross-family /
  multi-strategy comparison, multi-column tap selection #47), a local server, and the
  run/replay clock controls.
- **Parameter-space search strategies** (Bayesian/genetic) — pluggable policies
  atop the atomic sim unit (C12), not yet designed. Search over *params* sits atop the
  atomic unit; search over *structure* (graph mutation / crossover — NEAT-style)
  additionally needs topology-as-data (C24).
- **`aura new` scaffolder, the experiment-builder API, and `Aura.toml`'s
  static-context schema** — `aura new` scaffolds a Rust project *crate* (node /
  strategy / experiment blueprints) against the engine (C16/C20); the
  experiment-builder API surface (harness wiring, structural axes, sweep
  combinators) and `Aura.toml`'s schema (data paths, instrument/pip metadata,
  default broker & window, runs dir) are not yet designed.
- **The analysis meta-level — composable orchestration + project-as-crate authoring
  (tracked: #109).** aura's differentiator (C21) has two unbuilt halves: (1) the
  orchestration axes (sweep / Monte-Carlo / walk-forward) becoming *composable tools*
  a user wires into an analysis workflow, rather than today's hard-wired CLI verbs
  (`sweep_family` / `walkforward_family` in `aura-cli`); and (2) the project-as-crate
  authoring layer above (the `aura new` / experiment-builder API / `Aura.toml` thread
  + cdylib loading, C16/C17). On the meta-level, authoring a strategy is *wiring
  existing nodes* (C9 composition) + *defining the analysis framework* (C21), not
  writing new nodes; the user wants it driven via the CLI, later the interactive
  server / playground (C14/C22). **Resolved (2026-06-29) → C24.** The fork was
  mis-posed as "(a) run-a-Rust-experiment-via-CLI vs (b) wire-nodes-via-CLI". The real
  axis is **topology = compile-time Rust source vs runtime, serializable, World-owned
  data value**, settled as the **value** (C24, the game-engine principle): node
  *logic* stays Rust (C17), *topology* becomes data the World generates / serializes /
  reproduces. "Wire-nodes-via-CLI" is dropped — interactive wiring through the CLI is a
  non-goal; it collapses to a file + parser, which *is* C24's load path. What
  **remains** under #109: the **blueprint data format** itself (C24 Status — its own
  brainstorm / milestone), and the **composable-orchestration** half — the axes
  (sweep / MC / walk-forward) becoming tools applied to a World-constructed harness
  instead of the hard-wired `aura-cli` verbs (`sweep_family` / `walkforward_family`).
- **Inferential honesty of the World — family-selection without false-discovery
  control (tracked: milestone "Inferential validation (defend against false
  discovery at sweep scale)", #144 / #145 / #146; adjacent #139).** The World's
  massively-parallel sweep (C12 axes 1–2, C21) is aura's differentiator *and* its
  most direct route to false discovery: `optimize` / `rank_by` select the single
  best family member by a bare argmax, with no penalty for the number of
  configurations tried, no preference for a robust parameter neighbourhood over a
  sharp in-sample peak, and no cross-instrument generalization read; `mc` is a
  descriptive seed-sweep, not an inferential significance test. The hygiene
  invariants keep one backtest honest (C1 determinism, C2 no look-ahead); they do
  **not** make a *selection across a family* honest. Recognized direction: the
  selection/aggregation layer must deflate a winner for the family size (#144),
  prefer a plateau over a peak (#145), and score cross-instrument generalization
  (#146), with a per-candidate out-of-sample bootstrap as the adjacent
  significance read (#139, landed cycle 0075). Distinct from the two threads above —
  neither orchestration *composability* (#109) nor a search *policy* (Bayesian/
  genetic), but the statistical *validity* of the selection itself. Not a
  C-invariant until built; recorded as the World's third half — **now structurally
  complete** (all three pieces built, cycle 0078).
  **Status (cycle 0076):** the trials-deflation piece (#144) landed —
  `optimize_deflated` (aura-registry) wraps `optimize` and records, additively on
  the winning member's manifest (C18, the new `RunManifest.selection`), a deflated
  score + an overfit probability for the family size, **without changing which
  member wins** (C23; `optimize` / `rank_by` stay a bare argmax — the deflation is
  recorded provenance, not a re-ranking). The R arm is a centred moving-block
  reality-check (reusing the `r_bootstrap` kernel); the `total_pips` arm a
  closed-form expected-max-of-K dispersion floor. **Status (cycle 0077):** the
  plateau-over-peak piece (#145) landed — `optimize_plateau` (aura-registry)
  argmaxes the **neighbourhood-smoothed** metric surface (mean or worst-case over
  each member's closed mixed-radix grid neighbourhood) instead of the bare peak,
  recorded on the same `RunManifest.selection` carrier, now an orthogonal *rule ×
  annotation*: `SelectionMode::{Argmax, PlateauMean, PlateauWorst}` with either the
  deflation annotation (#144) or the plateau annotation (`neighbourhood_score` /
  `n_neighbours`). It is opt-in via a `--select` flag — default argmax stays
  byte-identical (C23). The grid lattice surfaces from the engine
  (`SweepBinder::sweep_with_lattice`); the policy stays in aura-registry with
  `walk_forward` selection-agnostic (C9). Selection is now a *pluggable objective*
  (bare argmax → trials-deflated → plateau), not a single hard-wired argmax.
  **Status (cycle 0078):** the last piece — cross-instrument generalization (#146) —
  landed, completing the inferential half. Unlike #144/#145 (which *select + annotate*
  a within-family winner on `RunManifest.selection`), #146 is an **aggregator/
  validator**, not a selector: the `aura generalize` subcommand runs a *brought*
  candidate across an instrument list and `generalization` (aura-registry) reduces the
  per-instrument R-metrics to a **worst-case floor** (min over instruments — the
  cross-instrument sibling of `PlateauMode::Worst`, R-only per C10) + a sign-agreement
  count + the per-instrument breakdown. It is a *recomputable aggregate* over a new
  `FamilyKind::CrossInstrument` family (C12's comparison axis, realized), each member
  self-identifying via the new `RunManifest.instrument` lineage field (C18) — *not*
  stamped on `FamilySelection` (there is no within-family winner to annotate). The
  inferential half is now structurally built; the per-candidate OOS bootstrap (#139,
  cycle 0075) is its adjacent significance read.
- **`aura-std` contents** — substantively populated (~30 node modules): SMA/EMA,
  arithmetic (`Add`/`Sub`/`Mul`/`Sqrt`/`LinComb`), logic (`And`/`Gt`/`Latch`/`EqConst`),
  the `Delay` z⁻¹ register, `Resample`, `RollingMin`/`RollingMax`, `Session`, the R chain
  (`Bias`/`FixedStop`/`Sizer`/`PositionManagement`), the legacy `SimBroker` pip yardstick,
  the cost-model graph (`CostNode`/`CostRunner`/`CostSum` + `ConstantCost`/`VolSlippageCost`/
  `CarryCost`), and the `Recorder`/`GatedRecorder`/`SeriesReducer` sinks. Further blocks
  land demand-driven as the walking-skeleton needs them.
- **`strategies/` split** — a later split, *inside a project*, of top-level
  strategies from reusable building blocks in `nodes/`; not a day-1 cut.
- **Sequencing** — the runnable single-harness *substrate* comes first (walking
  skeleton: a closed harness that runs deterministically and records via a sink);
  the World/meta layer (C21) and the playground trace-explorer (C22) are the
  differentiating layers that follow — you orchestrate and visualize a thing that
  must first run once.