Sibling of #40 (composite output as a named multi-field record). Brings
the other two composite-boundary edge-kinds to the same named-projection
shape: input roles (Vec<Role { name, targets }>, rendered [in:<name>])
and a param-alias overlay (Vec<ParamAlias { name, node, slot }>) that
relabels an interior leaf param slot's surface name in param_space().
Load-bearing decision recorded in the spec: param aliasing is a pure
naming overlay, NOT curation — every interior slot stays in param_space()
and sweepable; the alias only relabels in place, never reorders/hides
(forced by the issue's "legibility not capability" framing + C23). The
injected point vector is unchanged; MACD surfaces macd.fast/slow/signal
instead of three indistinguishable macd.length.
Gates green: feature-acceptance (worked MACD before->after is the
evidence), placeholder scan clean, parse-every-block no-op (profile
declares no spec_validation), grounding-check PASS (7 load-bearing
assumptions each ratified by a named currently-green test; compiled_view
+ param_space slot-order goldens stand as the C23 regression anchors).
refs #41
Composite::output: OutPort -> Vec<OutField { node, field, name }> so
multi-line indicators (MACD/Bollinger/Stochastic/Ichimoku) can be
authored as a composition and re-exported as a named record. Output
half only; param/input-role naming is the dependent sibling.
refs #40
Redesign the `aura graph` blueprint view (render_blueprint) from ascii-dag
subgraph cluster boxes to a "program with subroutines" model: a small main
graph that wires the harness with each composite shown as a single opaque
node, plus a `where:` section rendering each distinct composite type's
interior once. render_compilat (flat, fully-inlined, C23) is unchanged.
Design ratified in-session. The 3a-vs-3b fork was resolved by medium: the
interactive enter/focus model (3a) is the playground's, parked as #37; the
static main+definitions model (3b) is the CLI's, this cycle.
Substance:
- Correctness under width: ascii-dag 0.9.1's subgraph level-centering rounds
sibling x-positions with `/2`, overlapping wide sibling labels (a
width/parity-sensitive bug with no config/padding dodge that survives
unequal-width siblings). The flat layout is collision-free; the new model
renders only flat graphs, so the bug cannot arise.
- Scale: a composite is one opaque node in the main graph, so blueprint size
tracks top-level wiring, not inlined node count — real strategies stay
displayable.
- Render-once: a composite reused N times has its body rendered once.
- Removes the nested-composite `unimplemented!` at graph.rs:74-77 (the
definitions pass is recursive).
Defaults baked (not forks): composite-type dedup key = name() (assumes
same name => same structure, the authoring convention); opaque-node ports
shown as [in:k] / [out] markers in a definition.
grounding-check PASS — all nine load-bearing assumptions ratified by green
tests (Composite/Blueprint API, the flat render_compilat path, the
unimplemented! site, the two replaced cluster tests, the four must-stay-green
tests). spec_validation parse gate is a confirmed no-op (none configured).
refs #38
The amend chose `SMA(length)` (type + knob names) for the param-generic blueprint
view. Implementation surfaced that the ascii-dag 0.9.1 renderer writes a label
verbatim on one line (no multi-line — `render/ascii.rs::write_node` brackets the
raw string) and its Sugiyama subgraph layout overlaps two wide sibling boxes inside
a cluster (`[SMA(length[SMA(length)]`). Width is auto-computed but the cluster
packing does not honor it; the only spacing knob (`node_spacing`) is on the
deprecated config path, global, and width-independent — no robust option, and
domain labels grow unboundedly wide (`LinComb(weights[0], weights[1])`, deep
path-qualified names). Horizontal mode is already rejected (collapses fan-outs).
So `LeafFactory::label()` renders the bare node type (`SMA`). The tunable knobs are
surfaced by `param_space()`, not in the graph; the compiled view still labels built
nodes valued (`SMA(2)`) via `Node::label`. Correct C22 reading either way —
structure (now: type-only) before, values after.
refs #31
Plan-recon surfaced two surfaces the first draft under-scoped; resolved here after
a user design decision on the render form.
- Blueprint rendering (C22): a value-empty blueprint has no bound values, so the
`aura graph` blueprint view can no longer label a leaf `SMA(2)`. LeafFactory gains
`name` + `label()`, and the blueprint view renders the param-generic form
`SMA(length)` (type + tunable knob names, no values), chosen over a bare `SMA` to
surface the knobs the milestone is about. The compiled view (post-build flat
nodes) is unchanged — it still labels valued. Correct C22 reading: structure
before, values after.
- Vestigial pre-build interface removed: `Composite::schema` / `BlueprintNode::schema`
derive an item interface from built interior leaves; a value-empty leaf has none,
and both methods have no live caller (compile resolves every interface structurally
on the built flat nodes). Removed with their unit test. This is why the factory
carries no io skeleton — the build-then-wire decision stands, no schema duplication.
- OQ3: the swapped-sample mis-wiring moves from the builder args to the injected
vector; its observability moves to the compiled view (the blueprint view is
param-generic and identical for both orderings).
Grounding-check re-run: PASS (LeafFactory::label internally consistent; the
no-live-caller claim for the removed methods verified; blueprint-view vs
compiled-view goldens correctly partitioned).
refs #31
Cycle B of milestone "The World — parameter-space & sweep". A blueprint's tunable
values are injected at bootstrap rather than baked into the builder: a leaf becomes
a param-generic recipe (LeafFactory: params -> sized node), and a new build-then-
wire compile path binds a positional Scalar vector slot-by-slot against
param_space(), building each node through its own constructor.
Ratified design (brainstorm): value-empty B over mutate-in-place and over the
default-bearing variant. The value lives only in the injected vector (no baked
default), so the blueprint stays a pure param-generic recipe (C19); injected
values flow through the single new() sizing/validation gate; and because
compile_with_params consumes the vector in the same depth-first order param_space()
reports, the two share one traversal — subsuming #34's dual-traversal drift hazard.
Scope: one vector -> one instance; LeafFactory in aura-core; bootstrap_with_params/
compile_with_params; CompileError::{ParamKindMismatch, ParamArity}; Scalar::as_i64/
as_f64 accessors; fixtures/CLI/bit-identity re-expressed against the vector.
Deferred: sweep enumeration (#32), domain validation (#32/C20), single-run
authoring convenience (#35).
Gates: grounding-check PASS (8 current-behaviour assumptions ratified against green
tests; the sole BLOCK — a missing Scalar value accessor — resolved by adding it as
in-scope surface). Parse gate a no-op (profile declares no spec_validation).
refs #31
Cycle A of milestone 'The World — parameter-space & sweep'. A node declares
its tunable parameters in its C8 schema (ParamSpec { name, kind }, a third
NodeSchema field); Blueprint::param_space() aggregates them into one flat,
path-qualified param-space as a read-only projection of the graph-as-data.
Param identity is positional (slot after the deterministic inline order),
the name a non-load-bearing path-qualified debug symbol (C23). Vector knobs
(LinComb.weights) expand to N flat indexed entries; N is topology-fixed (C19).
compile/inline_composite are untouched — the compilat stays bit-identical.
Scope is declaration + aggregation + inspection only; binding (#31), sweep
enumeration (#32), search-range, constraint, and default-range are deferred.
refs #30
End-to-end milestone fieldtest (closing gate) for the "Construction layer"
milestone (#12 composites + #13 aura graph render). Four real downstream tasks
authored against the PUBLIC interface only (ledger + doc-comments + re-exports
+ the aura CLI), under fieldtests/milestone-construction-layer/:
mc_1 — author a named sma_cross composite, build a Blueprint, compile +
bootstrap + run (12 populated equity rows);
mc_2 — correct vs fast/slow-swapped cross: labels + graph-as-data differ
observably (the headline mis-wire-visible property);
mc_3 — composite-nested-in-composite inlines and runs;
mc_4 — walk the built graph via the read-only accessors, no engine internals.
Roll-up: friction_found, NO bugs. The milestone delivers its core promise —
fractal authoring -> compile -> bootstrap -> run, introspection as graph-as-data,
and param-carrying labels that make a mis-wire readable. None of the findings
block the gate.
Findings filed to the forward queue (not fixed here):
- #28 (feature) — the headline friction: `aura graph` renders only the built-in
sample and the render adapter is CLI-private, so a consumer can't render their
OWN graph. Subsumes the reachability of #26 (the nested-render panic is
structurally unreachable until the render is parameterizable). Likely next
(consumer-project / World) milestone.
- #29 (idea) — aura-engine doesn't re-export the scalar vocabulary (ScalarKind
etc.) a graph-builder needs; one-crate ergonomics tidy.
- #16 (comment) — `aura graph` arg policy (help / unknown-arg) should unify with
the still-open `aura run` strictness question.
The fieldtester read no implementation source; all artefacts are the consumer
crate + the two live render captures (render_clustered.txt / render_compiled.txt,
byte-identical across runs).
Render a wired graph as an ASCII DAG via an `aura graph` subcommand so a
mis-wiring becomes visible (the concern of #13). Two selectable views: a
blueprint view (composites as cluster boxes, pre-inline) and a `--compiled`
view (the flat compilat, boundaries dissolved per C23).
Settled in brainstorm: render labels come from a non-load-bearing `label()`
default method on the core `Node` trait (a C8 refinement aligned with C23,
which already reserves render names citing #13), overridden per node to carry
params (SMA(2) vs SMA(4)) so a swap reads back differently. Composites gain an
authored name; the engine exposes read-only graph-as-data accessors and stays
dependency-free (C16) — ascii-dag lives only in aura-cli. `aura run` is left
untouched this cycle.
Gates: self-review clean; parse gate a no-op (no spec_validation in profile);
grounding-check PASS.
refs #13
The Construction-layer milestone's anchor spec: a Blueprint graph-as-data that
compiles (inlines composites) to the flat (nodes, sources, edges) the unchanged
Harness::bootstrap consumes (C9/C19/C23). The SMA-cross composite runs
bit-identical to the hand-wired sample_harness (C1). Non-goals: optimisation
passes, external deps, named-handle ergonomics, aura graph.
Grounding-check PASS (13/13 current-behaviour assumptions ratified by green tests).
refs #12
Milestone-scope fieldtest (the functional leg of the Walking-skeleton
milestone-close gate), from the public interface only. Three end-to-end
scenarios, all green:
- the newcomer CLI path: `aura run` -> single-line {manifest,metrics} JSON,
exit 0, byte-identical twice (C1), non-degenerate trace (C22), bad-args ->
exit 2;
- the real-data thread: real AAPL.US 2006-08 M1 bars -> load_m1_window ->
close_stream -> SMA-cross -> Exposure -> SimBroker -> summarize -> RunReport,
deterministic AND populated;
- epoch-ns coherence: C3 normalization holds ingest -> run -> sink.
Roll-up: the milestone DELIVERS its promise end to end (ingest -> one signal ->
exposure -> sim-optimal broker -> pip-equity metric -> structured RunReport,
deterministic, populated) on BOTH the synthetic and real-data paths. 0 bugs.
Headline concern (verified by the orchestrator): issue #19's premise is WRONG at
the shipped SMA(2)/SMA(4) demo config — 21 AAPL.US bars warm the cross and
produce a populated trace (total_pips=-1.5, max_drawdown=65.5, 6 sign flips),
not all-zero. The degeneracy #19 saw was deep-SMA/param-dependent, not intrinsic
to the symbol. #19 to be re-scoped. Other findings: `aura run <junk>` exit 0
(already #16), `aura --help` lands on the exit-2 error path (friction), SimBroker
two-f64-slot order unchecked at bootstrap (documented footgun) — both filed.
Per-cycle fieldtest of the aura-ingest ingestion boundary (issue #7), from the
public interface only. 3 examples (transpose/normalize core; close_stream ->
deterministic harness run with epoch-ns end-to-end; load_m1_window over real
AAPL.US 2006-08 bars + None paths), all green. 0 bugs, 0 friction.
Two spec_gaps named for follow-up (neither a code defect; both filed to the
tracker):
- load_m1_window returns Some(empty M1Columns) — not None — for an in-coverage
but bar-empty window. The rustdoc's "None if no data in [from_ms,to_ms]" reads
bar-level but the realized boundary is file-level (it inherits data-server's
file-skip None). A consumer matching None == "no bars" is wrong for the
in-coverage-empty case. -> tighten the load_m1_window rustdoc (one line).
- AAPL.US (the carrier/integration-test symbol) has only ~21 M1 bars/month, so
the SMA-cross never warms and the end-to-end run yields all-zero (degenerate
but valid) metrics. The boundary is correct (C2 warm-up); the milestone's
shipped demo data is too thin to show a non-degenerate trace. -> weigh against
C22's "newcomer sees a populated trace" before the Walking-skeleton close
(document expected density or name a denser demo symbol/window).
First real data source (refs #7, Walking-skeleton milestone). A new
`aura-ingest` crate pulls data-server as a cargo git dependency, transposes
its AoS M1Parsed records into SoA base columns (C7), normalizes Unix-ms to
canonical epoch-ns at the one ingestion boundary (C3), and feeds the existing
k-way merge a real close-price stream so a backtest runs over real bars (C1).
Decisions captured:
- Eager materialization (not a lazy/shared Source abstraction): C12's cross-sim
Arc<[T]> sharing has no consumer until the multi-sim orchestration cycle, so
building it now would be speculative infra; deferred, transpose logic carries
over.
- aura-ingest is the external-dependency firewall: data-server transitively
pulls chrono+regex+zip; isolating it in its own crate keeps core/std/engine
zero-external-dep. cargo test --workspace now needs a populated cargo cache.
- Scope: boundary + tests only, no aura run CLI arg surface (would force the
unsettled #16 arg-parse decision); M1 first, tick deferred.
grounding-check returned BLOCK on the single load-bearing external assumption
(the data-server public API) — structurally unratifiable by an aura-side test
because data-server is brand-new this cycle; all 9 aura-internal assumptions
were ratified by named green tests. Overridden after verifying data-server's
API against its actual source (lib.rs DataServer::new/has_symbol/
stream_m1_windowed/next_chunk/DEFAULT_DATA_PATH; records.rs M1Parsed fields,
types, Copy), per the agent's documented override path. User-approved.
Public-interface-only field test of the walking-skeleton closing seam (#8):
the `aura run` CLI and the reusable aura-std::Recorder sink. A standalone
consumer crate (path-deps on the three engine crates, as a C16 project would)
drives the real binary as a subprocess and wires Recorder via the raw
Harness::bootstrap API.
Examples (all built from HEAD and ran green):
- c0010_1: drive `aura run` as a downstream CLI/jq consumer — single-line JSON
report, byte-identical across two runs (C1), bad-args → exit 2 + exact usage on
stderr with empty stdout.
- c0010_2: wire Recorder onto Sma(3) via raw bootstrap, drain the mpsc::Receiver
— rows match the rustdoc warm-up model exactly.
- c0010_3: Recorder over [I64, Bool, Timestamp] — the four-kind claim (never
exercised by the CLI's f64-only path) holds, verified from rustdoc alone.
Findings: 0 bugs, 4 working, 1 spec_gap, 1 friction.
- [spec_gap] `aura run <trailing-arg>` is accepted (exit 0), not rejected — the
arg parse keys only on the first token being `run` and ignores the rest. The
cycle_scope's "any other or missing subcommand -> exit 2" does not unambiguously
cover `run <junk>`; the binary chose the lenient reading. Tracked for a
ratify-or-tighten decision.
- [friction] verifying the documented {manifest, metrics} JSON nesting needs a
hand-rolled string walker — the zero-dep consumer has only to_json() and no
typed read-path. Low-priority tidy.
Both non-bug findings filed to the forward queue; neither blocks the cycle.
Cycle 0010, the Walking-skeleton milestone's closing seam: an `aura run`
subcommand that bootstraps a built-in sample harness (synthetic source →
SMA-cross → Exposure → SimBroker → recording sinks), runs it
deterministically, and prints the cycle-0009 metrics+manifest report as
canonical JSON to stdout.
The crux the cycle resolves: no recording sink node ships today (recording
lives only as #[cfg(test)] fixtures), and Harness::run returns () with a sink
as the only data-out path. So this cycle first ships a reusable
aura-std::Recorder (a pure consumer holding an mpsc::Sender, purity-preserving
per C7), then wires the CLI on top.
Load-bearing decisions, user-approved at the brainstorm gate:
- Recorder ships in aura-std (universal block, C16), not engine or CLI-local.
- Sample harness is authored as a plain Rust constructor over the raw
Harness::bootstrap API (C17/C20) — the open experiment-builder DSL thread is
deliberately NOT committed this cycle.
- Zero-dependency CLI: hand-parsed args (no clap); manifest commit from
option_env!("AURA_COMMIT") defaulting to "unknown" (build.rs git capture
deferred).
Grounding-check PASS (all load-bearing codebase assumptions ratified by green
tests). Non-goals: experiment-builder DSL, aura new, Aura.toml schema, the
data-server source (#7), build.rs git capture.
refs #8
First fieldtest of the run-metrics + manifest report surface. A standalone
downstream-consumer crate (fieldtests/cycle-0009-run-metrics/) path-depends on
the engine crates and exercises the post-0009 surface from the public interface
only (rustdoc + ledger + glossary + project layout, never crates/*/src).
Primary axis empirically met: the north-star "a run emits metrics + manifest"
move is reachable from rustdoc alone — drain two recording sinks -> f64_field ->
summarize -> RunManifest -> to_json, metrics matching the hand model on the first
run, deterministic across reruns.
Findings: 0 bugs, 1 friction, 1 spec_gap, 4 working.
- working x4: north-star reachable from rustdoc; SimBroker firing/slot docs (a
resolved 0007 gap) now carry the example; summarize metric definitions exact
on six degenerate inputs (incl. negative-curve drawdown + flat-as-sign-0);
f64_field panics precise and well-located.
- spec_gap: to_json's JSON key names + {manifest,metrics} nesting are not on
the public surface — a consumer parsing the JSON (C18 registry, the deferred
aura run printer) cannot author against it from rustdoc alone.
- friction: to_json renders whole-valued f64 without a decimal point (3.0 ->
"3"), so one f64 field appears as integer or decimal token within one schema.
Both doc-level findings are the same doc pass and matter mainly for the deferred
aura run (#8) and the C18 registry that will parse this JSON. Spec feeds the next
plan as reference.
refs #6
Cycle 0009 (Walking skeleton milestone, issue #6): a post-run pure-function
reduction over the recorded pip-equity + exposure streams into summary metrics
(total pips, peak-to-trough drawdown, exposure sign-flip count) plus a
caller-supplied RunManifest (commit, params, data-window, seed, broker), paired
as RunReport with canonical hand-rolled zero-dep JSON (C14). Lives in a new
aura-engine report module; no engine/Harness/node-contract change.
The reduction is post-run, not a node (C8 caps a node at one record per eval
with no end-of-run eval); matches C18's manifest+metrics-per-run-from-day-one.
The `aura run` subcommand that prints the report is deferred to #8.
Grounding-check PASS (all load-bearing assumptions ratified by green tests /
authoritative sources). Spec-validation parse pass is a documented no-op (no
parser configured for the profile).
refs #6
First fieldtest of the sum combinators. A standalone downstream-consumer crate
(fieldtests/cycle-0008-sum-combinators/) path-depends on the engine crates and
exercises the post-0008 surface from the public interface only (rustdoc + ledger
+ glossary + project-layout, never crates/*/src).
Primary axis empirically met: the north-star two-signal combine move now uses the
shipped Add (Add::new() dropped in exactly where the 0007 fixture hand-authored
an Add2) — the 0007 boilerplate is retired, end to end through SimBroker to a
deterministic recorded pip curve.
Findings: 0 bugs, 1 friction, 2 spec_gaps, 3 working.
- working ×3: Add2 retired; LinComb weighted/variadic sums exact (<1e-12);
warm-up barrier + empty-weights panic as documented.
- spec_gap: Add/LinComb per-input firing policy (Firing::Any / mode-A as-of
join) absent from the public surface — same class as the 0007 SimBroker gap.
- spec_gap: a fired combinator emits one row per *cycle*, so a heterogeneous
same-timestamp multi-source trace can carry >1 row per timestamp (correct per
C4/C5, undocumented at this surface).
- friction: nested consumer crate still needs the empty [workspace] table
(carried from 0006/0007; #9 closed the docs half; folds into aura new).
Spec feeds the next plan as reference.
Ships the missing sum combinator(s) in aura-std so the north-star combine move
(C10) is expressible from shipped blocks. Two nodes: Add (parameterless two-input
f64 sum, companion to Sub) and LinComb { weights } (N-input weighted sum, weights
as tunable params + arity, C8/C12). Both withhold output until all inputs are
present (no implicit cold-leg 0.0).
Resolves the cycle-0007 fieldtest [friction]: the consumer had to hand-author a
project-local Add2 to combine two signals.
refs #11
First fieldtest of the signal-quality loop. A standalone downstream-consumer crate
(fieldtests/cycle-0007-signal-quality/) path-depends on the engine crates and
exercises the post-0007 surface from the public interface only (rustdoc + specs +
ledger + glossary, never crates/*/src):
1 single-signal quality backtest, end to end (SMA-cross -> Exposure -> SimBroker
-> Recorder pip curve)
2 Exposure clamp + sizing (hard ±1 saturation, sign preserved)
3 sim-optimal integration: short-on-falling pays positive pips; pip_size=2 curve
exactly halves pip_size=1
4 north-star combine-two-signals (two MA-cross spreads summed into one exposure)
Findings: 3 working (carry-on), 2 spec_gap, 2 friction.
- spec_gap: SimBroker firing policy + cold-exposure->0.0 warm-up emission shape
not on the public surface (only in the feat commit body).
- spec_gap: SimBroker input slot order (0=exposure, 1=price) only in C10 prose /
commit body; a swapped wiring is not caught at bootstrap (both f64).
- friction: the north-star "combine two signals" move needs a sum/LinComb
combinator aura-std does not ship (hand-authored Add2 in the fixture).
- friction: standalone consumer crate still needs the empty [workspace] table
(carried from cycle-0006; tracked as #9).
Spec at docs/specs/fieldtest-0007-signal-quality.md feeds the next plan.
refs #4#5
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Cycle 0007 bundles Gitea #4 (exposure node) and #5 (sim-optimal broker) — the two
halves of the signal-quality loop, the realization of the C10 reframe. Ships two
aura-std nodes: Exposure { scale } (clamp(signal/scale, -1, +1)) and
SimBroker { pip_size } (causal prev_exposure * dprice / pip_size integration, no
look-ahead). Engine unchanged — pure node authoring on the frozen substrate.
grounding-check: PASS (all substrate assumptions ratified by green tests).
refs #4#5
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
First fieldtest of the project. A standalone downstream-consumer crate
(fieldtests/cycle-0006-substrate/) path-depends on the engine crates and
exercises the post-0006 substrate from the public interface only (rustdoc +
specs + ledger, never crates/ source):
1 custom recording node via the Node contract + Ctx::now()
2 fan-out/fan-in DAG, 3-arg bootstrap, run() -> ()
3 two interior streams recorded from one run (the cycle headline)
4 byte-identical recorded output across two runs (C1 determinism)
5 mis-wired recording edge rejected (KindMismatch) at bootstrap
Findings: 3 working (carry-on), 2 friction, 1 spec_gap.
- friction: a nested standalone consumer crate fights the workspace
resolver (needs an empty [workspace] table — Cargo's own hint).
- friction: recorder boilerplate is hand-rewritten per author (C9
deliberately ships no library sink).
- spec_gap: the public surface never states output: vec![] is THE sink
declaration, nor defines a Some return paired with empty output.
Spec at docs/specs/fieldtest-0006-substrate.md feeds the 0007 plan.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A sink is a role, not a type: recording is an out-of-graph side effect a
node performs in eval, pushing to a destination it holds as a field. No
Sink type, no trait, no engine sink-awareness. The engine stays the
router for in-graph routing (edge table = graph-as-data); the node-side
push is the only out-of-graph escape, and that boundary is the
determinism boundary. A node may be producer and sink at once (C8 both).
Engine surface shrinks: Ctx gains now(); the observe index (field,
bounds check, per-cycle collection) is removed; bootstrap drops its 4th
param; run returns (). Recording streams are sparse and timestamped
(only fired cycles), replacing the dense Vec<Option<row>>.
Realization notes land on C8 and C22; no new contract.
refs #2
Cycle 0005 (BLOCKER, milestone walking-skeleton): generalize a node's output
from a single scalar to a record of base-scalar columns (K >= 1), with a scalar
being the degenerate K = 1 case. A node keeps exactly one output port; that
port's payload becomes an ordered list of named, typed base columns (a composite
stream, C7). This unblocks every multi-column producer the engine needs next --
OHLCV bars and, later, the C10 position-event table -- none of which can exist
while a node emits at most one column. This revises C8 and sharpens C7.
The load-bearing design call, made with the user: the eval output representation
under C7's "no per-event heap allocation". eval becomes
`-> Option<&[Scalar]>` -- the node fills a buffer it owns (sized once at
construction) and returns a borrowed K-field row; the engine copies it into one
reused scratch row (resolving the borrow, zero per-cycle alloc once warm) and
scatters each field to its out-edges. `None` still means filter/not-warmed.
Rejected alternatives, on substance: returning `Option<Vec<Scalar>>` allocates
per fire (violates C7's forbids); an inline fixed `[Scalar; N]` record bakes a
magic width cap; engine-owned output columns invert the eval model (write-into-out
+ a fired flag) and split "fired" from "produced" for an output-lookback benefit
we do not need (history is bound as an input, not read from a node's own past).
The borrowed row is the minimal faithful generalization of today's
`Some(Scalar)` push route, K-wide.
Binding stays field-wise only: `Edge` gains `from_field`; 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. The
record is structural bundling over the four base types (C7 intact) -- no fifth
scalar type, no dyn Any, no heterogeneous buffer.
Scope is the substrate only: the contract change, the Sma/Sub degenerate
migration, the engine from_field wiring, and a neutral OHLCV proof (a 5-field
bundler fed by five timestamp-aligned barrier sources + a downstream Sub binding
high - low field-wise, plus must-fail cases: from_field out of range -> BadIndex,
per-field kind mismatch -> KindMismatch). The C10 position table and broker nodes
are out of scope -- later cycles that consume this capability.
Self-review clean; grounding-check PASS (14 load-bearing "before"-state
assumptions ratified by named green tests; C6/C7/C8 match the ledger as written;
parse-every-block a documented no-op -- profile declares no spec_validation).
refs #1
refs walking-skeleton
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fourth walking-skeleton cycle: the reactive-semantics layer — both firing
policies (C6), the k-way ingestion merge (C3), the monotonic cycle_id clock
(C4), and freshness-gated recompute / sample-and-hold (C5). This is the first
point where the engine departs from RustAst rather than sharpening it.
Both rails ship together (user directive: "must sit right from the start"):
- A = fire-on-any-fresh + hold (as-of join);
- B = all-fresh barrier (synchronizing join), accumulating across cycles.
Per-input-group, mixable on one node. Encoded as Firing::{Any, Barrier(u8)} on
InputSpec (where C8 places firing); default Any keeps Sma/Sub unchanged.
Key design resolution (studied against RustAst src/ast/rtl/streams/): the
barrier token is the data *timestamp*, not cycle_id. RustAst's
last_cycle_per_input.all(== cycle_id) synchronizes one source's fan-out but
cannot synchronize independent sources — under C4 four same-timestamp sources
are four different cycle_ids, so the cycle_id barrier never fires for C6's own
OHLC example. Substituting timestamp is the minimal faithful generalization
(fires both the diamond rejoin and the multi-source bar). Mode A and the k-way
merge are both new (RustAst had neither).
Two read clocks, both load-bearing (resolves the 0003 audit's "0004 owns
cycle_id" with no write-only field): cycle_id is the freshness epoch (C5 reads
it via per-slot fresh_at == cycle_id); timestamp is the barrier token (C6 reads
it via last_ts == T). Sample-and-hold falls out of the existing push model — a
non-firing node pushes nothing, so consumers see the held value via window[0].
Scope (user-chosen): reactive semantics only — in-memory (timestamp, value)
sources; the merge is the permanent ingestion core, the Source trait (C11) +
data-server IO are a later orthogonal cycle. bootstrap/run/InputSpec signatures
change (the 0003 tests re-express on the multi-source API).
Headline test pair (identical sources, firing-only difference): A emits
[None,None,120,130,140,240] (holds); B emits [None,None,120,None,None,240]
(waits for timestamp coincidence). Plus determinism (C1), mixed A+B, the 0003
fan-out/join compat baseline, and the bootstrap rejections.
Grounding-check PASS: 10 load-bearing substrate assumptions ratified by green
tests; the spec's "before" shapes match disk exactly.
refs walking-skeleton
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Third walking-skeleton cycle: aura-engine's first real content — a Sim that
runs a wired DAG of nodes deterministically, cycle by cycle. First point where
authored nodes execute against data, not just under a hand-fed Ctx.
Engine shape (contract-forced): a flat Vec<NodeBox> + an index edge table,
topologically ordered (Kahn), owned &mut by the loop — the flat, monomorphized
sharpening of RustAst's Rc<RefCell<dyn Observer>> push graph (C1/C7). Each node
owns its input columns (cycle-0002 shape), so Ctx is unchanged and the per-cycle
path allocates nothing. Wiring is by forwarding: a producer's eval output is
pushed into its consumers' input columns; a None forwards nothing (the seed of
sample-and-hold).
Bootstrap generalizes C7's "type check paid once at wiring" to the whole
topology: size input columns from each node's schema, kind-check every edge,
reject directed cycles (BootstrapError). A worked 2-input Sub node lands in
aura-std so the loop is proven on a real fan-out+join DAG
(source -> {SMA(2), SMA(4)} -> Sub -> observe), with a determinism assertion (C1).
Deliberately deferred (recorded as decisions): freshness gating / sample-and-hold
(C5 -> cycle 0004, only testable once a second source at a different rate exists);
the Source trait + data-server ingestion + k-way merge (C3/C11); the ergonomic
builder API (C19); the real sink + run registry (C18/C22); cross-sim parallelism.
Grounding-check PASS: all assumptions about the 0001 substrate and 0002 node
contract ratified by green tests; no phantom methods.
refs walking-skeleton
The second walking-skeleton cycle, on top of the 0001 streaming substrate:
the Node contract (C8) and its evaluation context.
Scope (one iteration):
- `Node` trait — `schema() -> NodeSchema` (inputs: kind + lookback; output
kind) + `eval(&mut self, Ctx) -> Option<Scalar>` (None = filter / not warmed).
- `Ctx` — a Copy borrow-wrapper handing eval zero-copy, financial-indexed
Windows per input (`ctx.f64_in(i)[k]`, index 0 = newest).
- Read-side `AnyColumn::as_f64/as_i64/as_bool/as_ts` — mirrors the existing
write-side `as_*_mut` and closes the cycle-0001 audit gap.
- `Sma` in aura-std — the worked producer node (the skeleton's first block),
driven by a hand-written test that mimics the future sim loop.
Deliberate deferrals (recorded as decisions, not gaps): the sim loop (C4),
freshness gating (C5), firing policies (C6, so InputSpec carries no firing
field yet), schema-level tunable params (C12/C19), and sinks / no-output
nodes (C8 consumer side). The hand-driven test stands in for the loop.
Grounding-check PASS: all load-bearing assumptions about the 0001 substrate
(Column/Window/AnyColumn/Scalar, financial indexing) ratified by green tests.
refs walking-skeleton
The first walking-skeleton cycle: the aura-core data substrate everything
streams through — the four scalar base types (Scalar/ScalarKind/Timestamp,
C7), the fixed-capacity financial-indexed SoA Column with its zero-copy
Window (C8), the type-erased AnyColumn edge with an edge-time kind check
(C7), and the per-column run_count freshness primitive (C5). Sharpens the
RustAst reference: no boxed Value, no VecDeque, no RefCell/Rc, no per-event
allocation. Node trait, Ctx, firing policies, and the sim loop are deferred
to subsequent cycles.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Cargo workspace aura-core -> aura-engine -> aura-cli (bin `aura`) plus nodes/ for hot-reloadable cdylib node crates. Crate bodies are intentionally API-free; types arrive from the first spec. Adds the skills profile (.claude/dev-cycle-profile.yml), the project CLAUDE.md with the eight domain invariants, and the design-ledger skeleton.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Brummel