aa94f51d8a997f6cd5055c42035826de4e5d0f75
13 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 claude
|
b048923f1c |
feat(engine): shared rayon pool for the disjoint-parallel core
Replace the per-call std::thread::scope fan-out in run_indexed with one process-wide rayon pool. run_indexed's callers nest — sweep (grid points), monte_carlo (seeds), and walk_forward (window bounds), where a walk_forward window runs an inner sweep — and each std::thread::scope spawned its own available_parallelism() OS-thread batch, so the two levels multiplied to ~available_parallelism()^2 threads (measured 150-240 concurrent on a 24-core box during a 42-cell campaign, ~10x oversubscription). rayon's pool is process-wide and persistent: a nested par_iter work-steals within the same N workers instead of spawning, so live workers never exceed the pool size. Determinism (C1) is preserved bit-for-bit: run_indexed collects via IndexedParallelIterator in enumeration order (not completion order), and the float aggregation (summarize_r / McAggregate / stitch) runs after the ordered collect, so IEEE-754 non-associativity introduces no variance. The existing _with_threads(1) == _with_threads(8) == public determinism tests stay green, now driving local rayon pools of 1 and N workers via ThreadPool::install. Structural inversion: the public sweep/monte_carlo/walk_forward become the core (calling run_indexed on the ambient pool through a shared assemble_* helper); the _with_threads variants collapse into thin, test-only wrappers (#[cfg(test)]) that run the same assembly inside a local pool. run_one stays Fn + Sync (no + Send): run_indexed borrows it in the map closure and the wrappers borrow it into install — passing it by value would move it into rayon's Map and force F: Send, a bound the public callers do not carry (hence the load-bearing #[allow(clippy::redundant_closure)] on run_indexed). Adds a nested-oversubscription regression test (asserts nested run_indexed never exceeds the pool size — the pre-rayon shape ran POOL^2) and an ignored wall-time benchmark. rayon admitted under the amended-C16 per-case dependency policy; it stays within C1 (parallelism across sims, never within one). This is the shared pool only — the first of three steps toward saturating the cores. The diagnosis found the larger loss is the idle valleys (~42% of the box idle: the sequential cell loop and the single-threaded bootstrap), which the pool alone does not fill. Two follow-ups are filed and recorded on #268: the Registry write path is not concurrency-safe (append_family races on a shared families.jsonl), the prerequisite for parallelizing the C1-disjoint cell loop with a RAM budget bounding concurrently-active instruments (the FileCache retains a symbol's parsed files while any reference is live). Verified: cargo test -p aura-engine (276 passed, the C1 determinism tests green), the new oversubscription test green, the benchmark runs (no overflow), clippy -D warnings clean, cargo test --workspace no failures. closes #268 |
||
|
claude
|
1ebb94c1b8 |
feat(engine,cli): run manifests stamp untouched bound defaults (closes #249)
RunManifest gains defaults: Vec<(String, Scalar)> — the wrap-prefixed bound_param_space() of the signal, read after axis reopening, so a bound param an axis overrode has already left the space and flows through params instead (disjoint by construction; verified end to end: a sweep member's overridden fast.length sits in params while slow.length/bias.scale sit in defaults). params keeps its "what varied" semantics and stays the reproduce input. One-directional serde widening (#[serde(default)]) per the selection/instrument/topology_hash idiom — old records deserialize with an empty defaults; unlike the Option fields it always serializes, mirroring params. ~20 struct-literal sites across five crates gained the field (compile-mandated breadth, no behaviour change at those sites). The C14 ledger records the underlying decision (2026-07-13): generated outcome records spend redundancy on direct readability (single writer, cannot drift); authored intent artifacts admit none (every redundancy is a drift site) — so the fix lands in the manifest, never the blueprint. Verification: RED test run_manifest_stamps_untouched_bound_defaults green; cargo build --workspace; cargo test --workspace green; clippy -D warnings on the touched crates; binary-level sweep exclusivity check. |
||
 Brummel
|
4928e289f7 |
feat(project): the project-as-crate load boundary (cycle 0102)
A research project is now a loadable external cdylib crate. Inside a directory whose ancestry holds an Aura.toml, aura discovers the project root cargo-style, locates the compiled dylib via cargo metadata (debug default, --release opt-in), loads it load-and-hold, and refuses mismatches before trusting anything: the AURA_PROJECT descriptor (aura-core::project, #[repr(C)]) carries a C-ABI stamp prefix (rustc + aura-core version, baked per consuming build by the new aura-core build.rs) validated before any Rust-ABI field is read. The vocabulary charter gates the merged resolution: project type ids are ::-namespaced (std stays bare), duplicates refuse, and the enumerable type-id list must agree with the resolver, so introspection can never silently omit a project type. All blueprint verbs resolve through the merged project + std vocabulary via a per-invocation Env threaded through the dispatch chains; registry, trace-store, and data paths anchor at the project runs root (Aura.toml [paths], paths-only by design — instrument geometry stays the recorded sidecar, C15). RunManifest gains the Tier-1 project provenance field (namespace + dylib sha256 + best-effort commit), stamped beside topology_hash on the blueprint-run paths; pre-0102 registry lines load unchanged. Default node names strip the namespace, so :: never reaches the param-path address space. Proven by the demo-project fixture (built by the e2e via cargo, path-dep on this workspace): run twice bit-identical, provenance recorded, introspection lists demo::* beside std, registry anchors at the discovered root from a subdirectory; the badcharter fixture proves the charter refusal through the real libloading path; a never-built project refuses with a cargo-build hint. Outside a project every path collapses to the previous literals — goldens and manifest pins byte-identical. Verification: cargo build --workspace clean; cargo test --workspace 862 passed / 0 failed (incl. 7 project_load e2e); clippy -D warnings clean (one precedent-matching allow(too_many_arguments) on run_oos_blueprint, whose arity the Env threading raised to 8); doc build unchanged. Docs/ledger aligned: Aura.toml field lists are paths-only in project-layout.md, glossary, C16/C17; new C13 realization note records the per-invocation-reload reading and the load-and-hold one-shot scope boundary. New deps, per-case review (aura-cli leaf binary only, never the frozen artifact): libloading, toml. refs #180 |
||
|
Brummel
|
ff4a1b3d4a |
feat(0092): run-from-blueprint infrastructure — restructure + topology_hash + shared seam (Tasks 1-3)
Cycle-1 infrastructure for #165 (World/C21). Behaviour-preserving + workspace green; the CLI arm + tests (Tasks 4-5) follow. Task 1 — restructured stage1_r_graph into stage1_signal() (the serializable signal leg: SMA-cross -> Bias, a `price` input-role + a `bias` output) + wrap_stage1r(signal, ...) (broker/sinks/exec/cost around a nested signal), with stage1_r_graph() = wrap_stage1r(stage1_signal(...)). DEVIATION from the plan's "callers untouched" claim, verified behaviour-preserving: nesting the signal prefixes its param-space names (stage1_signal.fast.length), which broke the sweep's bare .axis("fast.length"); fixed exactly as the #137 stop-knob machinery does — FAST_LENGTH_SUFFIX/SLOW_LENGTH_SUFFIX suffix-resolution + stage1_r_friendly_name arms mapping the prefixed names back to the bare manifest names. All observable behaviour (manifest names, member keys, metrics, traces) byte-identical; the flat graph and every metric unchanged (full suite green, incl. sweep/walkforward/MC). Task 2 — RunManifest.topology_hash: Option<String> (Tier-1 optional, serde default/skip — old manifests byte-identical, #156), threaded None into all 24 literal sites across 7 crates. Also fixed the aura-registry RunManifestRead read-mirror to carry topology_hash (it was hardcoding None on load) — else a stored hash would silently drop on the registry round-trip once Task 3 stores a real one. Task 3 — the shared run_signal_stage1r seam (hash the signal, wrap, compile with params, bootstrap, run, manifest with params from param_space + topology_hash) + the sha256_hex topology_hash helper (sha2 in aura-cli, off the frozen engine, invariant 8). run_stage1_r now carries its topology_hash too (every run becomes self-identifying, #158). The seam is unwired until Task 4 (transient #[allow(dead_code)], retired there); RED-first seam tests added. Verified: cargo test --workspace green (51 suites, 0 failures); cargo clippy --workspace --all-targets -D warnings clean. refs #165 |
||
|
Brummel
|
0a06afed29 |
feat(0078): cross-instrument generalization as a validation step
Last piece of the Inferential-validation milestone (#146): a new `aura generalize` subcommand grades how consistently a single candidate holds across a set of instruments — the across-instrument axis of the same anti-false-discovery concern as trials-deflation (#144, within one family) and plateau-over-peak (#145, within one instrument's surface). Aggregator/validator, not a third selector (user-ratified): it runs a brought candidate (a single-cell stage1-r grid — --fast/--slow/--stop- length/--stop-k pinned to single values) across an instrument list (--real SYM1,SYM2,...), collects per-instrument R-metrics, and reduces them to a worst-case floor (min over instruments) + a sign-agreement count + the per-instrument breakdown. R-only (C10 — R is the only account/instrument-agnostic unit); the worst-case min is the cross- instrument sibling of PlateauMode::Worst and satisfies the non- domination constraint by construction (a strong instrument cannot lift a min). Rejected: a cross-instrument t-stat mean/std*sqrt(M), which reintroduces the pooled mean the milestone warns against. - RunManifest gains a first-class `instrument` lineage field (serde- widened, C14/C18 — legacy lines load as None, existing manifest bytes unchanged); the compat read-mirror threads it in lockstep. - FamilyKind::CrossInstrument (C12 comparison axis); the M per-instrument runs persist as a CrossInstrument family, each member self-identifying via its stamped instrument. The aggregate is recomputable from the members (McAggregate precedent), printed live. - The generalization reduction lives registry-side beside optimize_ deflated/optimize_plateau (C9), reusing resolve_metric/metric_value/ is_r_metric; check_r_metric is the data-free pre-check the CLI runs before evaluating any instrument. - Additive: existing sweep/walkforward/mc/standalone paths stay byte- identical (C23 — instrument omitted when None via skip_serializing_if). Verified by the orchestrator: cargo test --workspace green (0 failed — the new generalization_*, check_r_metric, parse_generalize_*, and the gated cross-instrument E2E all pass; the E2E ran the real GER40/USDJPY path on this host, exit 0); cargo clippy --workspace --all-targets -D warnings clean. closes #146 |
||
|
Brummel
|
a2959050a4 |
feat(0076): deflate the sweep winner's metric for the number of trials
Adds optimize_deflated beside optimize in aura-registry: the in-sample walk-forward selection now records, on each OOS winner's manifest, how much the winner's metric is inflated by the size of the search it won — a deflated score and (R arm) an overfit probability — without changing which member wins (additive, C23; a regression test pins optimize_deflated's winner == optimize's). - R arm (sqn_normalized/expectancy_r/...): a centred moving-block reality-check. Each member's per-trade R series is mean-subtracted to impose the no-edge null, resampled via the r_bootstrap kernel (extracted to a shared resample_block, byte-identical), and the best-of-K null-max gives overfit_probability = (count(null >= raw)+1)/(n+1) and deflated_score = raw - p95(null). Deterministic given the seed (C1). - total_pips arm: a closed-form expected-max-of-K dispersion floor (inv_norm_cdf + expected_max_of_normals), no probability. The record (FamilySelection on RunManifest) rides the proven serde(default, skip_serializing_if) widening + a compat.rs mirror, so legacy runs.jsonl/families.jsonl lines load unchanged and unstamped sweep/mc/standalone manifests stay byte-identical (C14/C18). The SelectionMode enum reserves a Plateau slot for #145. CLI: walk-forward stamps both arms; runs family <id> rank surfaces a human-readable deflated/overfit line. Quality review (re-dispatched after the implement-loop's quality gate exhausted) caught a real defect: on the R arm with positive resamples but no member carrying trade_rs (a family loaded from disk — trade_rs is serde-skipped), the null was all-NEG_INFINITY, turning deflated_score into +inf. Fixed by collapsing "no usable null" (zero resamples OR no trade_rs) into one degenerate floor (deflated == raw, overfit == 1.0); pinned by optimize_deflated_no_trade_rs_floors_instead_of_infinity. Verified: cargo test --workspace and cargo clippy --workspace --all-targets -D warnings both green. Frictionless Stage-1 R; n_eff advisory and reload-time recompute deferred (spec 0076 Out of scope). closes #144 |
||
|
Brummel
|
68605b7d88 |
refactor(stage1-r): rename the strategy-output param namespace exposure -> bias
Complete the deferred rename tail from cycle 0065's #126 (which renamed only the typed metric, keeping the param namespace uniformly "exposure" so the single-run rename would not smuggle in a ~30-site sweep-pinned change). Now renamed together as one consistent unit, all driven by the Bias node's instance name: - Bias::builder().named("exposure") -> .named("bias") (the harness builders + the engine fixtures in test_fixtures.rs / blueprint.rs); - the derived knob path exposure.scale -> bias.scale (Composite::param_space derives the knob from the node name) -- the sweep .axis/.range bindings and .with() points (blueprint.rs, main.rs), the walkforward ParamSpecs (walkforward.rs), the member_key dir names, and the JSON byte-pins in cli_run.rs / report.rs / main.rs tests; - the exposure_scale manifest param label -> bias_scale (the single-run manifest builders + the streaming_seam / real_bars fixtures). The rename is surgical: the three targets (named("exposure"), exposure.scale, exposure_scale) are syntactically distinct from the deliberately-kept "exposure" forms, which are untouched -- SimBroker's pre-reframe exposure input slot + prev_exposure + the exposure-integral; the LongOnly / gate / latch ports named "exposure"; and the on-disk "exposure" tap label (the recorded bias series that feeds `chart --tap exposure`, decoupled from the node name via ColumnarTrace::from_rows("exposure", ...)). No on-disk back-compat break: the knob path is a runtime param address, and recorded params in old runs.jsonl / families.jsonl are inert data strings (a pre-existing recorded "exposure.scale" still loads, it is just data). INDEX.md cycle-0065 realization note amended: the param namespace is recorded as the now-completed tail; the SimBroker slot + the on-disk tap label remain the deliberate permanent keeps (#117), and exposure_sign_flips stays a serde alias. Verified: cargo build --workspace --all-targets clean; clippy --all-targets -D warnings clean; full suite 514 passed / 0 failed (source + byte-pins renamed together); live `aura run` emits "bias_scale", `aura sweep` emits "bias.scale". closes #134 refs #126 #117 |
||
|
Brummel
|
d858caf67b |
chore: scrub dangling references to deleted specs/plans from sources
docs/specs and docs/plans were retired (prior commit); the source/test
comments that cited them ("spec 0050 §4.1", "spec §Testing N", "per spec",
"the spec's ...") now point at nothing. Strip every such pointer while
preserving the technical substance, the design-ledger contract refs
(C1/C11/C20/C34/C12.1/...), and the Gitea issue refs (#41).
Comment/doc edits only across 20 files — no logic change; full workspace
suite green, clippy clean.
|
||
|
Brummel
|
8b22fa6c73 |
feat(aura-engine): WalkForwardResult::named_params — mirror SweepFamily::named_params
Add the typed/named convenience view on `WalkForwardResult`, closing the symmetry cycle 0048 left open: it shares the `(space, tag-free Vec<Cell>)` idiom with `SweepFamily` but had no named-params accessor, so a consumer hand-wrote `zip_params(&result.space, &result.windows[i].run.chosen_params)`. `named_params(window)` pairs each param-space name with that window's chosen value in slot order, delegating to `zip_params` — the direct mirror of `SweepFamily::named_params`. A derived view, no new stored state. closes #76 |
||
|
Brummel
|
9febdb9623 |
refactor(aura-engine): walk-forward param plane — space on family, tag-free chosen_params, schema-checked param_stability
Make the walk-forward param plane C7-clean and symmetric with the sweep param plane. WalkForwardResult gains `space: Vec<ParamSpec>` (the kinds, once, mirroring SweepFamily.space) and WindowRun.chosen_params changes Vec<Scalar> -> Vec<Cell> (the tag-free coordinate the inner sweep already produces). param_stability now resolves its numeric coercion once per slot against the schema (result.space) into a Vec<fn(Cell)->f64>, then runs a type-blind reduction loop — replacing the old per-window-value scalar_as_f64 with its per-value unreachable!. scalar_as_f64 is deleted.
This overturns cycle 0047's "Fork A" (chosen_params stays Vec<Scalar>), which was justified by Scalar's new serializability. That defense was moot: WalkForwardResult and WindowRun carry no serde derives — they are never serialized; only the embedded oos_report (a RunReport, typed via the 0047 lift) is. So chosen_params is a pure in-memory substrate for param_stability, and Vec<Scalar> made it C7-impure (carrying the kind N×M times at the value, where the kind is per-slot constant). The substantive case — kind at the schema, SweepFamily symmetry, per-value unreachable! eliminated — wins, the same argument-form that retired the {kind,cell} Scalar struct in 0047.
The seam ([A]): walk_forward / walk_forward_with_threads take `space: Vec<ParamSpec>` by value, a sibling of the run-window closure; the Fn(WindowBounds)->WindowRun+Sync bound is unchanged and run_indexed is untouched (space is never captured by the Sync closure, moved onto the result at the end). A debug_assert in walk_forward_with_threads guards the arity coupling (every window's chosen_params.len() == space.len(), guaranteed by the same blueprint). walkforward_family computes the space once before the roll and passes the tag-free winner (chosen_params: best.params) directly, dropping the from_cell+zip reconstruction.
[D]: a Bool param slot coerces to 0/1 (mean = fraction "on"), keeping the coercer total over the knob kinds; a Timestamp slot is structurally impossible (C20) so the schema pass carries one unreachable! (once per slot, never per value). param_stability keeps an `if result.windows.is_empty() { return Vec::new(); }` guard to preserve its documented empty-on-no-windows contract. WalkForwardResult and SweepFamily share the (space + tag-free Cell points + zip_params) idiom but stay distinct types — different axis (Sweep enumerates input coordinates; WF rolls time and chosen_params is the optimizer's output), and WF additionally carries bounds + oos_equity + stitched_oos_equity.
Behaviour-preserving (C1): the three migrated walk-forward tests stay green with identical MetricStats (param_stability_reduces_chosen_params_per_slot still pins mean 2.5 / p50 2.5 / mean 1.5); the i64->f64 and f64 coercions are value-identical to the deleted scalar_as_f64. The one genuinely new behaviour (Bool 0/1) is unexercised by built-in grids and pinned by the new param_stability_reduces_bool_slot_to_fraction_true (mean 0.75, 3-of-4 true).
Gates: build --workspace --all-targets, test --workspace, clippy --workspace --all-targets -D warnings, doc --workspace --no-deps — all clean.
closes #75
|
||
|
Brummel
|
86746e3d5d |
refactor(aura-core): Scalar as a native tagged enum, disjoint from Cell; typed RunManifest.params
Scalar was `struct { kind: ScalarKind, cell: Cell }` — "a Cell wearing a kind hat." The recorded reason for that shape was migration ease, which is not a design rationale (CLAUDE.md: rationale != effort), so the struct had no substantive defense. Redefine it as the native tagged union it conceptually is:
enum Scalar { I64(i64), F64(f64), Bool(bool), Timestamp(Timestamp) }
This is substantively better on four axes: kind/bits skew becomes unrepresentable (illegal states gone); accessors panic loudly on the wrong variant instead of a release-mode silent bit reinterpret; PartialEq derives the documented IEEE-754 / cross-kind value semantics (the hand-roll, needed only because the struct inherited Cell's bitwise compare, is gone); and serde is a plain derive emitting the externally-tagged wire form ({"I64":10}/{"F64":2.5}) — the private ScalarRepr shadow enum that motivated this was never needed. It is also C7-honest: erased-on-the-hot-path (Cell) and self-describing-at-the-edge (Scalar) are two disjoint types bridged by explicit conversion.
The whole public API is preserved (Scalar's fields were private), so call sites do not churn: the enum change is contained to scalar.rs, where cell() now encodes and from_cell() decodes per kind. Cell and ScalarKind are untouched.
With Scalar serializable, lift RunManifest.params from Vec<(String, f64)> to Vec<(String, Scalar)>: the param's kind (an i64 length vs an f64 scale) now survives into the C18 record (runs.jsonl) and the CLI JSON instead of collapsing to f64. scalar_as_param_f64 is deleted; sim_optimal_manifest passes typed params through. This is a deliberate wire-shape change — params now render as tagged scalars; the JSON-asserting tests are updated to the new shape on purpose.
Hand-authored manifest params across the CLI's single-run/mc/macd sites use honest kinds (lengths -> i64, scales -> f64) so they match the sweep path (which already derives correct kinds via zip_params); their in-binary JSON assertions are re-tagged accordingly.
Walk-forward fork resolves with no code change: WindowRun.chosen_params stays Vec<Scalar> (the serializable record carrier), reduced to f64 only at the param_stability statistic boundary (scalar_as_f64 retained). Glossary 'cell' entry updated to describe Scalar as the disjoint tagged union, not 'a cell plus its kind tag'.
Gates: build --all-targets, test --workspace (incl. new scalar_serde_round_trips), clippy -D warnings, doc --no-deps — all clean.
|
||
|
Brummel
|
cd3d1ca9ed |
refactor(aura-core): split Scalar into a tag-free Cell + ScalarKind
Motivation
----------
`Scalar` was a tagged enum (I64/F64/Bool/Ts), so every scalar value
physically carried its own kind tag. But the kind is already known from
the schema/port/column the value flows through (C7: the type is a
property of the column, not of the value — the hot path is already
columnar `Column<T>`, and `AnyColumn::get` *reconstructs* the tag from
the column on the way out). The per-value tag was therefore redundant
with the kind the surrounding context already holds.
That redundancy had three costs:
* It baked an implicit `match` (a branch) into every function that read
a Scalar payload — even where the caller statically knew the type.
The tag could never be exploited away.
* Size: a tagged enum is tag + payload = 16 bytes (f64/i64 alignment),
twice the 8 bytes the value needs. A `Column<Scalar>` would be double
the memory and half the cache utilisation.
* It is the shared root of several downstream papercuts we keep hitting
— the lossy f64 manifest field, the `unreachable!` panic on a
non-numeric param, the serde-tag question — all symptoms of "the type
is baked into the value".
Change
------
Introduce `Cell`: a type-erased 64-bit word (`struct Cell(u64)`) that is
not readable without external type context. It is constructed per base
type (`from_i64/from_f64/from_bool/from_ts`) and read only by naming the
type at the call site (`i64()/f64()/bool()/ts()`) — each a branch-free
bit-cast. The hot path resolves the kind once at the boundary (from the
schema) and then reads natively, with no per-value branch. `Cell` knows
nothing of `Scalar` or `ScalarKind`; the dependency is strictly one-way,
and it lives in its own `cell.rs` (more is planned on top of it).
`Scalar` becomes `struct { kind: ScalarKind, cell: Cell }` — the
self-describing form for the dynamic boundaries (builder binding,
serialization, rendering), built on top of `Cell`. Its `as_*` accessors
now `debug_assert` the kind and return the native value (free in
release); calling the wrong accessor is a caller bug, not a checked
`Option`. The variant constructors `Scalar::I64(..)` become associated
fns `Scalar::i64(..)`.
`PartialEq` is hand-written (not derived) to preserve the former enum's
value semantics: kinds must match, then native payloads compare, so f64
keeps IEEE-754 behaviour (`NaN != NaN`, `+0.0 == -0.0`) and a kind
mismatch is never equal even when the raw words coincide. A fixture
(`scalar_eq_is_value_not_bitwise`) pins exactly the cases where bit- and
value-equality diverge, so it can't silently regress. `Cell`'s own
`Eq`/`Hash` stay bitwise — correct for a raw word.
The change is behaviour-preserving: Scalar's observable behaviour is
identical to the pre-Cell enum (the value-equality fixture proves it);
only the internal representation changed. The ~440 call sites across the
workspace are a mechanical constructor rename plus ~12 destructuring
sites (match-arms / `let`-patterns) rewritten to `kind()` + `as_*`.
Verified: cargo build --workspace --all-targets, cargo clippy --workspace
--all-targets -- -D warnings, cargo test --workspace — all green.
|
||
|
Brummel
|
4764656062 |
feat(aura-engine): walk-forward family — WindowRoller + walk_forward (C12 axis 3)
The third C12 orchestration axis: walk-forward varies the data window. A
WindowRoller is a pure iterator of WindowBounds { is, oos } — bounds only, never
tick data (#71 firewall). walk_forward runs a caller closure per split disjointly
over the shared run_indexed core (C1), and stitches the OOS pip-equity segments
into one continuous curve (each segment offset by the running sum of prior
segments' finals; an empty segment contributes 0.0, mirroring summarize). C2
no-look-ahead is structural: the roller only ever emits oos.0 == is.1 + 1, pinned
by a zero-tick bounds test. The in-sample optimize (axis 2) is closure-supplied,
not called by the engine: aura-engine has no aura-registry dependency (C9), and
C12 forbids baking search policy into the primitive — the CLI bridges both crates.
Param-stability is on-demand, not stored (the R2 decision, recorded with
provenance on #69): WalkForwardResult stores only the raw per-window outcomes +
the stitched curve, and param_stability(&result) -> Vec<MetricStats> is a public
reduction over the retained per-window chosen params. A stored summary would be
recomputable from the windows (redundant) and would force one statistic canonical
when "stability" admits several; this mirrors SweepFamily (raw points + on-demand
optimize), not McFamily (stored aggregate). MetricStats::from_values is extracted
from McAggregate::from_draws (behaviour-preserving — the 7 mc tests stay green)
and gains serde so the CLI summary renders it and #70 lineage can persist it.
`aura walkforward` runs a built-in rolling walk-forward over a synthetic windowed
source: per window it sweeps the built-in grid in-sample, optimizes by total_pips,
runs the chosen params out-of-sample, persists each OOS RunReport (C18), and prints
a stitched summary line.
Two plan deviations, both compiler-forced and consistent with siblings: added
#[derive(Debug)] to WindowRoller (the RED tests' unwrap_err needs Ok: Debug; the
sibling types already derive Debug); removed a now-redundant test-module
SyntheticSpec import after Step 2 hoisted it to the top-level use.
Verification: cargo test --workspace green (aura-engine 152 incl. walkforward 9 +
mc 9, aura-cli 14); clippy --workspace --all-targets -D warnings exit 0; cargo doc
--workspace --no-deps clean.
refs #69
|