Files
Aura/crates/aura-campaign/src/exec.rs
T
claude d3b1a1aead feat(campaign,registry,cli): per-cell fault isolation — a failed cell is recorded, never a global abort
closes #272

A member fault (no-data, bind, run, or a caught panic) is now a recorded
per-cell outcome instead of aborting the whole campaign and discarding every
already-computed cell. The incident that motivated this (a 22-instrument
campaign lost ~36 healthy cells ~6.7 min in because Copper had an archive gap)
now completes: the healthy cells persist, the gap cell is recorded as failed,
and the run exits 3.

Direction (owner decision 2026-07-14): run to completion and report
compromised results; no coverage preflight, no window synthesis.

Containment granularity:
- The CELL for a sweep-stage member fault (a grid hole structurally
  compromises winner selection, so the whole cell fails).
- The FOLD for a walk_forward member fault (independent time windows): the
  surviving folds pool into the family, failed folds are recorded as
  StageRealization.window_faults, and the summary names the ratio.

- aura-registry: additive CellFault / CellFaultKind (closed:
  no_data|bind|run|panic|window) / WindowFault / CellCoverage, plus
  fault/coverage fields on CellRealization and window_faults on
  StageRealization — all serde-default-skipped, so pre-#272 campaign_runs
  lines parse and round-trip byte-identical.
- aura-campaign: run_cell returns a fault-annotated CellRealization instead of
  Err (execute's accumulate-then-append-once tail is unchanged and now
  persists every healthy cell + the one run record); a `contain` split keeps
  ExecFault::Registry and doc-shape preflight faults global while Member/Window
  become per-cell/per-fold. Member panics are caught with
  catch_unwind(AssertUnwindSafe) at all three member-run sites (sweep IS/OOS)
  and recorded as MemberFault::Panic — a member panic no longer aborts the
  process. The wf stage partitions Registry faults (global) from Member/Window
  (per-fold) and filters faulted-fold placeholders (the
  "faulted-member-placeholder" broker sentinel) out of the persisted family.
- aura-cli: exec_fault_prose gains the Panic arm; CliMemberRunner::window_coverage
  derives effective bounds + interior gap months from the #264 archive
  primitives; present_campaign prints per-cell failure notes + a completion
  summary and threads the failed-cell count; a run with >=1 failed cell exits 3
  ("completed with failed cells") uniformly across `aura campaign run` and the
  dissolved sweep/walkforward/mc/generalize verbs (exit_on_campaign_result).
  Usage stays 2, refused-before-running stays 1, clean stays 0.

Tests: the global-abort pins flip to containment (execute + the two wf fault
tests → fold-containment + all-folds-fail-the-cell); new panic-containment
tests on both the sweep path (PanicRunner) and the wf path (this commit adds
the wf mirror the loop left uncovered); a new gapped-archive e2e (one covered
cell + one gap cell → exit 3); the ~14 CLI exit-1 pins move to the exit-3
register; a pre-#272-line byte-identical round-trip guard.

Suite: cargo test --workspace green (1309 tests, 0 failed); clippy clean.
Decision log: #272 comments (fork rationale, the fold Registry/Member split,
the placeholder sentinel, uniform exit-3).

Follow-up (minor, not blocking): the plan under-scoped Task 1 to aura-registry
though the additive fields also touch aura-campaign's exec.rs literals — the
loop absorbed it mechanically; a future plan for a cross-crate additive-field
change should scope every crate's construction sites in the first task.
2026-07-14 16:51:31 +02:00

1132 lines
50 KiB
Rust

//! Campaign execution (#198): the cell loop over strategy x instrument x
//! window, the v1 stage semantics (sweep members -> family -> selection;
//! per-member gate filtering; the walk-forward seam), and the realization
//! record. Precondition: the CALLER has run the doc tiers (intrinsic +
//! referential validation), so this module re-checks nothing a validated
//! document guarantees (e.g. non-empty axes); execution-level refusals live
//! in the crate's preflight, which `execute` runs before any member runs.
use std::collections::BTreeMap;
use std::sync::Mutex;
use aura_analysis::{FamilySelection, SelectionMode};
use aura_core::{zip_params, Cell, ParamSpec, Scalar, Timestamp};
use aura_engine::{
r_bootstrap, sweep, walk_forward, GridSpace, ListSpace, RollMode, RunManifest, RunMetrics,
RunReport, Space, SweepFamily, SweepPoint, WalkForwardError, WindowBounds, WindowRoller,
WindowRun,
};
use aura_registry::{
derive_trace_name, generalization, optimize, optimize_deflated, optimize_plateau,
sweep_member_reports, CampaignGeneralization, CampaignRunRecord, CellRealization, FamilyKind,
PlateauMode, Registry, StageBootstrap, StageRealization, StageSelection,
};
use aura_research::{Axis, CampaignDoc, DocRef, ProcessDoc, SelectRule, StageBlock, WfMode};
use crate::{
member_metric, predicate_holds, preflight, CellSpec, ExecFault, MemberFault, MemberRunner,
DEFLATION_BLOCK_LEN, DEFLATION_N_RESAMPLES,
};
/// One family-producing stage's full member payload — what an emit renderer
/// (`family_table`) prints, in ordinal order.
#[derive(Clone, Debug)]
pub struct StageFamily {
pub stage: usize,
pub block: &'static str,
pub family_id: String,
pub reports: Vec<RunReport>,
}
/// One selection-bearing stage's payload — what an emit renderer
/// (`selection_report`) prints.
#[derive(Clone, Debug)]
pub struct StageSelectionOut {
pub stage: usize,
pub block: &'static str,
pub family_id: String,
pub winner_ordinal: usize,
pub params: Vec<(String, Scalar)>,
pub selection: FamilySelection,
}
/// Per-cell stage payloads, for the consumer's emit rendering.
#[derive(Clone, Debug)]
pub struct CellOutcome {
pub families: Vec<StageFamily>,
pub selections: Vec<StageSelectionOut>,
/// The cell's nominated generalize candidate (spec-fixed): the LAST
/// walk-forward window's OOS report + chosen params when a wf stage ran
/// (the freshest out-of-sample evidence for the finally-selected params),
/// else the sweep stage's winner; `None` for gate-truncated cells.
pub nominee: Option<(Vec<(String, Scalar)>, RunReport)>,
}
/// Everything one campaign run produced: the stored realization record, the
/// assigned run counter, and the per-cell payloads.
#[derive(Clone, Debug)]
pub struct CampaignOutcome {
pub record: CampaignRunRecord,
pub run: usize,
pub cells: Vec<CellOutcome>,
}
/// Execute a validated campaign document's process pipeline once per
/// (strategy, instrument, window) cell, in doc order, sequentially (C1:
/// parallelism lives inside the engine `sweep`, across members, never across
/// cells). `campaign_id` is the campaign document's 64-hex content id (its
/// first 8 chars prefix family names); `strategies` is index-aligned with
/// `campaign.strategies` as (blueprint content id, canonical blueprint json).
/// Writes one family per family-producing stage plus one campaign-run record;
/// a `MemberFault` aborts the whole run before any write of the faulted stage.
pub fn execute(
campaign_id: &str,
campaign: &CampaignDoc,
process: &ProcessDoc,
strategies: &[(String, String)],
runner: &dyn MemberRunner,
registry: &Registry,
) -> Result<CampaignOutcome, ExecFault> {
preflight(process, campaign)?;
if strategies.len() != campaign.strategies.len() {
return Err(ExecFault::PipelineShape {
detail: format!(
"resolved strategies ({}) are not index-aligned with campaign.strategies ({})",
strategies.len(),
campaign.strategies.len(),
),
});
}
let process_id = match &campaign.process.r#ref {
DocRef::ContentId(id) | DocRef::IdentityId(id) => id.clone(),
};
// The store's self-keying always yields a 64-hex content id; anything else
// is a caller bug — refuse it typed instead of panicking on the prefix
// slice below.
if campaign_id.len() != 64
|| !campaign_id.bytes().all(|b| matches!(b, b'0'..=b'9' | b'a'..=b'f'))
{
return Err(ExecFault::PipelineShape {
detail: format!(
"campaign_id must be a 64-hex content id (got {} chars)",
campaign_id.len()
),
});
}
let campaign_prefix = &campaign_id[..8];
let mut cells_out: Vec<CellOutcome> = Vec::new();
let mut cells_rec: Vec<CellRealization> = Vec::new();
// Per-(strategy_ordinal, window_ordinal) nominees across instruments, in
// cell-loop (instrument) order — the campaign-scope generalize input.
type Nominee = Option<(Vec<(String, Scalar)>, RunReport)>;
let mut nominees: BTreeMap<(usize, usize, usize), Vec<(String, Nominee)>> = BTreeMap::new();
// Absent/empty risk = a single default cell (regime None, ordinal 0); a
// non-empty list maps each regime to Some. The default's value is the member
// runner's — the doc is never mutated (its content id hashes with risk absent).
let regimes: Vec<(usize, Option<aura_research::RiskRegime>)> = if campaign.risk.is_empty() {
vec![(0, None)]
} else {
campaign.risk.iter().copied().enumerate().map(|(i, r)| (i, Some(r))).collect()
};
for (strategy_ordinal, (entry, (strategy_id, blueprint_json))) in
campaign.strategies.iter().zip(strategies).enumerate()
{
for instrument in &campaign.data.instruments {
for (window_ordinal, window) in campaign.data.windows.iter().enumerate() {
for (regime_ordinal, regime) in &regimes {
let cell = CellSpec {
strategy_ordinal,
strategy_id: strategy_id.clone(),
blueprint_json: blueprint_json.clone(),
axes: entry.axes.clone(),
instrument: instrument.clone(),
window_ms: (window.from_ms, window.to_ms),
regime: *regime,
regime_ordinal: *regime_ordinal,
};
let (outcome, realization) = run_cell(
&cell,
process,
campaign_prefix,
window_ordinal,
campaign.seed,
runner,
registry,
)?;
nominees
.entry((strategy_ordinal, window_ordinal, *regime_ordinal))
.or_default()
.push((instrument.clone(), outcome.nominee.clone()));
cells_out.push(outcome);
cells_rec.push(realization);
}
}
}
}
// Campaign-scope generalize (#200 d2): for each (strategy, window), grade
// the per-instrument nominees via the shipped `generalization`; fewer than
// 2 winners is a recorded shortfall, never computed around.
let generalize_metric = process.pipeline.iter().find_map(|s| match s {
StageBlock::Generalize { metric } => Some(metric.clone()),
_ => None,
});
let mut generalizations: Vec<CampaignGeneralization> = Vec::new();
if let Some(metric) = generalize_metric {
for ((strategy_ordinal, window_ordinal, regime_ordinal), cells) in &nominees {
let winners: Vec<(String, Vec<(String, Scalar)>)> = cells
.iter()
.filter_map(|(inst, nom)| {
nom.as_ref().map(|(params, _)| (inst.clone(), params.clone()))
})
.collect();
let missing: Vec<String> = cells
.iter()
.filter(|(_, nom)| nom.is_none())
.map(|(inst, _)| inst.clone())
.collect();
let pairs: Vec<(String, &RunReport)> = cells
.iter()
.filter_map(|(inst, nom)| nom.as_ref().map(|(_, report)| (inst.clone(), report)))
.collect();
let grade = if pairs.len() >= 2 {
Some(generalization(&pairs, &metric).map_err(ExecFault::Registry)?)
} else {
None
};
generalizations.push(CampaignGeneralization {
strategy_ordinal: *strategy_ordinal,
window_ordinal: *window_ordinal,
regime_ordinal: *regime_ordinal,
generalization: grade,
winners,
missing,
});
}
}
let mut record = CampaignRunRecord {
campaign: campaign_id.to_string(),
process: process_id,
run: 0,
seed: campaign.seed,
cells: cells_rec,
generalizations,
// The claim sentinel (#201 d5): a `Some` (any content) tells
// `append_campaign_run` to compose the derived trace name onto the
// stored line; `None` means the document requests no taps.
trace_name: (!campaign.presentation.persist_taps.is_empty()).then(String::new),
};
let run = registry.append_campaign_run(&record).map_err(ExecFault::Registry)?;
// Mirror the store-side derivation onto the returned copy (the 0109
// contract: stored line and returned record carry the same
// `Some("{campaign8}-{run}")` iff persist_taps is non-empty, else None) —
// keyed off the claim already stamped above, via the registry's own
// `derive_trace_name` (the one source of truth for the composition), not
// a re-evaluation of `persist_taps`.
let trace_name = record.trace_name.is_some().then(|| derive_trace_name(campaign_id, run));
record.run = run;
record.trace_name = trace_name;
Ok(CampaignOutcome { record, run, cells: cells_out })
}
/// The inter-stage population crossing `run_cell`'s seam: parameter points
/// only (from `std::grid`) or executed members (from `std::sweep`). Preflight
/// guarantees report-consuming stages (`std::gate`, `std::monte_carlo`'s
/// per-survivor arm) never see `Points`; the defensive faults below keep that
/// invariant loud if the shape rules ever drift.
enum StageFlow {
/// (ordinal into the nearest preceding family_id-bearing stage's family,
/// param point, member report) — the shape preflight established before
/// std::sweep populated it.
Members(Vec<(usize, Vec<Scalar>, RunReport)>),
Points(Vec<Vec<Scalar>>),
}
impl StageFlow {
/// Every flow can cross the seam as bare points — the walk_forward need.
fn points(&self) -> Vec<Vec<Scalar>> {
match self {
StageFlow::Points(points) => points.clone(),
StageFlow::Members(members) => {
members.iter().map(|(_, point, _)| point.clone()).collect()
}
}
}
}
/// Run one cell's pipeline: stages in doc order, a realized prefix that stops
/// at an empty gate (decision 8 — the cell truncates, the campaign continues).
fn run_cell(
cell: &CellSpec,
process: &ProcessDoc,
campaign_prefix: &str,
window_ordinal: usize,
seed: u64,
runner: &dyn MemberRunner,
registry: &Registry,
) -> Result<(CellOutcome, CellRealization), ExecFault> {
let grid = enumerate_grid(&cell.axes);
let mut families: Vec<StageFamily> = Vec::new();
let mut selections: Vec<StageSelectionOut> = Vec::new();
let mut stages: Vec<StageRealization> = Vec::new();
// The surviving population crossing the stage seam (see `StageFlow`).
let mut flow = StageFlow::Members(Vec::new());
// The cell's nominated generalize candidate: last wf window's OOS report
// when a walk_forward ran, else the sweep winner; None when truncated.
let mut nominee: Option<(Vec<(String, Scalar)>, RunReport)> = None;
for (stage_ordinal, stage) in process.pipeline.iter().enumerate() {
match stage {
StageBlock::Grid => {
// Enumerate-only (#256 fork B): the grid's parameter points
// cross the seam; nothing executes, nothing persists.
flow = StageFlow::Points(grid.points.clone());
stages.push(StageRealization {
block: "std::grid".to_string(),
family_id: None,
survivor_ordinals: None,
selection: None,
bootstrap: None,
window_faults: vec![],
});
}
StageBlock::Sweep { selection } => {
// Deterministic, self-describing family name (the "-{run}"
// suffix is appended by the registry).
let family_name = format!(
"{campaign_prefix}-{}-{}-w{window_ordinal}-r{}-s{stage_ordinal}",
cell.strategy_ordinal, cell.instrument, cell.regime_ordinal,
);
let family = match run_members(cell, &grid, runner) {
Ok(family) => family,
Err(fault) => {
// Registry etc. still aborts (contain's Err arm).
let cf = contain(fault, stage_ordinal)?;
return Ok(failed_cell(cell, stages, families, selections, cf, runner));
}
};
let family_id = registry
.append_family(&family_name, FamilyKind::Sweep, &sweep_member_reports(&family))
.map_err(ExecFault::Registry)?;
match selection {
Some(sel) => {
let (winner, selection) = select_sweep_winner(
&family,
&grid.axis_lens,
&sel.metric,
sel.select,
sel.deflate,
seed,
)?;
let winner_ordinal = family
.points
.iter()
.position(|p| p == &winner)
.expect("the winner is a member of its own family");
let params = zip_params(&family.space, &winner.params);
selections.push(StageSelectionOut {
stage: stage_ordinal,
block: "std::sweep",
family_id: family_id.clone(),
winner_ordinal,
params: params.clone(),
selection: selection.clone(),
});
// Nominate the sweep winner (superseded by a later
// walk_forward stage; cleared by an empty gate).
nominee = Some((params.clone(), winner.report.clone()));
stages.push(StageRealization {
block: "std::sweep".to_string(),
family_id: Some(family_id.clone()),
survivor_ordinals: None,
selection: Some(StageSelection { winner_ordinal, params, selection }),
bootstrap: None,
window_faults: vec![],
});
}
None => {
// Selection-free sweep (terminal, preflight-checked):
// the family is the whole result — no StageSelection
// is recorded (recording one would fabricate intent),
// no nominee is produced.
stages.push(StageRealization {
block: "std::sweep".to_string(),
family_id: Some(family_id.clone()),
survivor_ordinals: None,
selection: None,
bootstrap: None,
window_faults: vec![],
});
}
}
// The whole family flows on (decision 3: `select` names the
// recorded selection, never a filter).
flow = StageFlow::Members(
family
.points
.iter()
.enumerate()
.map(|(i, p)| (i, scalar_point(&family.space, &p.params), p.report.clone()))
.collect(),
);
families.push(StageFamily {
stage: stage_ordinal,
block: "std::sweep",
family_id,
reports: family.points.into_iter().map(|p| p.report).collect(),
});
}
StageBlock::Gate { all } => {
// A member survives iff ALL predicates hold; the crate-root
// `predicate_holds` compares the resolved per-member metric.
// An unresolvable metric (an R name against `metrics.r ==
// None`) fails the member — conservative and deterministic
// (the metric NAME was already preflighted against
// `PER_MEMBER_METRICS`).
let StageFlow::Members(members) = &mut flow else {
return Err(ExecFault::PipelineShape {
detail: format!(
"stage {stage_ordinal}: std::gate needs executed \
members (preflight admits no std::grid predecessor)"
),
});
};
members.retain(|(_, _, report)| {
all.iter().all(|p| {
member_metric(report, &p.metric)
.is_some_and(|value| predicate_holds(&p.cmp, value, p.value))
})
});
let ordinals: Vec<usize> = members.iter().map(|(i, _, _)| *i).collect();
let empty = ordinals.is_empty();
stages.push(StageRealization {
block: "std::gate".to_string(),
family_id: None,
survivor_ordinals: Some(ordinals),
selection: None,
bootstrap: None,
window_faults: vec![],
});
if empty {
nominee = None; // a truncated cell nominates no candidate
break; // decision 8: truncate this cell, keep running cells
}
}
StageBlock::WalkForward { .. } => {
// Deterministic, self-describing family name (the "-{run}"
// suffix is appended by the registry).
let family_name = format!(
"{campaign_prefix}-{}-{}-w{window_ordinal}-r{}-s{stage_ordinal}",
cell.strategy_ordinal, cell.instrument, cell.regime_ordinal,
);
// The seam crossing is plain points: the walk-forward stage
// re-sweeps the surviving param points and never needs the
// gate bookkeeping's ordinals or reports.
let survivor_points: Vec<Vec<Scalar>> = flow.points();
match run_walk_forward_stage(
seed,
cell,
stage_ordinal,
stage,
&grid.specs,
&grid.axis_lens,
&survivor_points,
&family_name,
runner,
registry,
) {
Ok(WfStageOutcome::Ran { family: fam, window_faults }) => {
stages.push(StageRealization {
block: "std::walk_forward".to_string(),
family_id: Some(fam.family_id.clone()),
survivor_ordinals: None,
selection: None,
bootstrap: None,
window_faults,
});
// Nominate the LAST window's OOS report (roll order,
// over the surviving folds) with its chosen params —
// the wf stage stamps them on the fresh OOS report's
// manifest.
nominee = fam
.reports
.last()
.map(|last| (last.manifest.params.clone(), last.clone()));
families.push(fam);
}
Ok(WfStageOutcome::CellFailed(cf)) => {
return Ok(failed_cell(cell, stages, families, selections, cf, runner));
}
Err(e) => return Err(e), // Registry / global refusal
}
}
StageBlock::MonteCarlo { resamples, block_len } => {
// Terminal annotator (#200 d1/d3): nothing flows out. Dual
// input: after a walk_forward, ONE bootstrap over the pooled
// per-window OOS trade-R series (roll order); after
// sweep/gates, one bootstrap per surviving member. Seeded from
// the campaign seed (the deflation convention, C1).
let bootstrap =
match families.iter().rev().find(|f| f.block == "std::walk_forward") {
Some(fam) => StageBootstrap::PooledOos(r_bootstrap(
&pooled_net_trade_rs(&fam.reports),
*resamples as usize,
*block_len as usize,
seed,
)),
// Zero-trade members (r: None) feed an empty slice —
// the engine's defined all-zero degenerate, recorded
// explicitly (a null result is a valid result).
None => {
let StageFlow::Members(members) = &flow else {
return Err(ExecFault::PipelineShape {
detail: format!(
"stage {stage_ordinal}: std::monte_carlo \
per-survivor bootstrap needs executed members \
(preflight admits no std::grid predecessor)"
),
});
};
StageBootstrap::PerSurvivor(
members
.iter()
.map(|(ordinal, _, report)| {
let rs = report
.metrics
.r
.as_ref()
.map(|r| r.net_trade_rs.as_slice())
.unwrap_or(&[]);
(
*ordinal,
r_bootstrap(
rs,
*resamples as usize,
*block_len as usize,
seed,
),
)
})
.collect(),
)
}
};
stages.push(StageRealization {
block: "std::monte_carlo".to_string(),
family_id: None,
survivor_ordinals: None,
selection: None,
bootstrap: Some(bootstrap),
window_faults: vec![],
});
}
StageBlock::Generalize { .. } => {
// Campaign-scope stage (#200 d2): `execute` runs it AFTER the
// cell loop, over the per-cell nominees across instruments —
// nothing realizes per cell.
}
}
}
Ok((
CellOutcome { families, selections, nominee },
CellRealization {
strategy: cell.strategy_id.clone(),
instrument: cell.instrument.clone(),
window_ms: cell.window_ms,
regime: cell.regime,
regime_ordinal: cell.regime_ordinal,
fault: None,
coverage: runner.window_coverage(cell),
stages,
},
))
}
/// The enumerated sweep grid of one cell: param specs (axis name = key, kind =
/// axis kind) in BTreeMap (lexicographic) key order, the per-axis radixes, and
/// the odometer point list — LAST axis fastest, the `GridSpace` order
/// discipline (`optimize_plateau` reads `axis_lens` in exactly this order).
struct SweepGrid {
specs: Vec<ParamSpec>,
axis_lens: Vec<usize>,
points: Vec<Vec<Scalar>>,
}
/// Odometer enumeration over the axes map, delegated to the engine's own
/// `GridSpace` (last axis fastest) rather than a hand-synced second copy of
/// the odometer — single-sourcing the order `optimize_plateau` reads
/// `axis_lens` in. Zero axes yield the single empty point. Empty VALUE lists
/// cannot reach here (`validate_campaign` refuses `EmptyAxis` upstream),
/// hence the `expect`.
fn enumerate_grid(axes: &BTreeMap<String, Axis>) -> SweepGrid {
let specs: Vec<ParamSpec> = axes
.iter()
.map(|(name, axis)| ParamSpec { name: name.clone(), kind: axis.kind })
.collect();
let values: Vec<Vec<Scalar>> = axes.values().map(|axis| axis.values.clone()).collect();
let grid = GridSpace::new(&specs, values)
.expect("axes validated upstream (validate_campaign refuses EmptyAxis)");
let axis_lens = grid.axis_lens();
let points: Vec<Vec<Scalar>> = grid
.points()
.into_iter()
.map(|cells| cells.iter().zip(&specs).map(|(c, ps)| Scalar::from_cell(ps.kind, *c)).collect())
.collect();
SweepGrid { specs, axis_lens, points }
}
/// Ref-counted panic-hook silencer: while at least one `SilencedPanic` guard
/// is alive, the process's panic hook is a no-op, so a member panic CONTAINED
/// by one of this module's three `catch_unwind` sites (recorded as a
/// `MemberFault`/`WindowFault`, the run continuing past it) does not also
/// paint a crash-looking default backtrace to stderr — "recorded, campaign
/// continues" stays observably true, not just true in the record. Ref-counted
/// rather than a bare set/restore because `sweep`/`walk_forward` run these
/// closures on multiple OS threads concurrently (C1 disjoint-parallel
/// members): a naive per-closure set_hook/take_hook pair would race two
/// threads' install/restore against each other. The previous hook (whatever
/// it was — default or a caller's own) is saved on the 0→1 transition and
/// restored on the 1→0 transition, so this never permanently clobbers a
/// hook installed above this module.
struct SilencedPanic;
#[allow(clippy::type_complexity)]
static PANIC_HOOK_STATE: Mutex<(usize, Option<Box<dyn Fn(&std::panic::PanicHookInfo<'_>) + Sync + Send>>)> =
Mutex::new((0, None));
impl SilencedPanic {
/// Enter the silenced section; drop the returned guard to leave it. Hold
/// this only around the `catch_unwind` call itself, never the whole
/// member-run closure, so a genuinely uncontained escape (this module's
/// own bug, not the member's) still surfaces normally once the guard has
/// dropped and unwinding continues past `catch_unwind`.
fn enter() -> Self {
let mut state = PANIC_HOOK_STATE.lock().expect("panic hook state lock");
if state.0 == 0 {
state.1 = Some(std::panic::take_hook());
std::panic::set_hook(Box::new(|_| {}));
}
state.0 += 1;
SilencedPanic
}
}
impl Drop for SilencedPanic {
fn drop(&mut self) {
let mut state = PANIC_HOOK_STATE.lock().expect("panic hook state lock");
state.0 -= 1;
if state.0 == 0
&& let Some(prev) = state.1.take()
{
std::panic::set_hook(prev);
}
}
}
/// Run every grid point through the consumer's `MemberRunner` via the engine
/// `sweep` (disjoint parallel members, C1 enumeration order). Faults are
/// captured per slot (the closure must return a report), and the LOWEST
/// enumeration index's fault aborts the stage after the sweep joins —
/// deterministic regardless of thread completion order. The placeholder
/// report for a faulted slot is never persisted or ranked: a captured fault
/// returns before any family write.
fn run_members(
cell: &CellSpec,
grid: &SweepGrid,
runner: &dyn MemberRunner,
) -> Result<SweepFamily, ExecFault> {
let space = ListSpace::new(&grid.specs, grid.points.clone()).map_err(|e| {
ExecFault::PipelineShape {
detail: format!("axis grid does not form a valid param space: {e:?}"),
}
})?;
// The engine closure receives only the point's cells; recover the
// enumeration index by value (duplicate points — possible only when an
// axis repeats a value — collapse to the first index, which is exactly
// the lowest-index attribution this capture exists for).
let cells: Vec<Vec<Cell>> =
grid.points.iter().map(|p| p.iter().map(|s| s.cell()).collect()).collect();
let faults: Mutex<Vec<(usize, MemberFault)>> = Mutex::new(Vec::new());
let family = sweep(&space, |point: &[Cell]| {
let params = zip_params(&grid.specs, point);
let called = {
let _silence = SilencedPanic::enter();
std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
runner.run_member(cell, &params, cell.window_ms)
}))
};
match called {
Ok(Ok(report)) => report,
Ok(Err(fault)) => {
let index = cells
.iter()
.position(|c| c.as_slice() == point)
.expect("the sweep only enumerates the grid's own points");
faults.lock().expect("fault capture lock").push((index, fault));
placeholder_report(&params, cell.window_ms)
}
Err(payload) => {
let index = cells
.iter()
.position(|c| c.as_slice() == point)
.expect("the sweep only enumerates the grid's own points");
faults
.lock()
.expect("fault capture lock")
.push((index, MemberFault::Panic(panic_message(payload))));
placeholder_report(&params, cell.window_ms)
}
}
});
let mut captured = faults.into_inner().expect("fault capture lock");
captured.sort_by_key(|&(index, _)| index);
if let Some((_, fault)) = captured.into_iter().next() {
return Err(ExecFault::Member(fault));
}
Ok(family)
}
/// #272: which executor faults are per-cell outcomes vs campaign-wide
/// refusals. Member and window faults are data-and-cell-local; everything else
/// (doc-shape, registry) stays a global refusal returned to the caller.
fn contain(fault: ExecFault, stage: usize) -> Result<aura_registry::CellFault, ExecFault> {
match fault {
ExecFault::Member(m) => Ok(aura_registry::CellFault {
stage,
kind: member_kind(&m),
detail: member_fault_prose(&m),
}),
ExecFault::Window { stage: s, detail } => Ok(aura_registry::CellFault {
stage: s,
kind: aura_registry::CellFaultKind::Window,
detail,
}),
other => Err(other),
}
}
fn member_kind(m: &MemberFault) -> aura_registry::CellFaultKind {
use aura_registry::CellFaultKind as K;
match m {
MemberFault::NoData { .. } => K::NoData,
MemberFault::Bind(_) => K::Bind,
MemberFault::Run(_) => K::Run,
MemberFault::Panic(_) => K::Panic,
}
}
/// Prose for a member fault. `pub` (re-exported at the crate root) so the
/// CLI's `exec_fault_prose` phrases the `ExecFault::Member` arm by calling
/// this directly instead of carrying its own duplicate of the wording.
/// Debug-leak-free, self-describing.
pub fn member_fault_prose(m: &MemberFault) -> String {
match m {
MemberFault::NoData { instrument, window_ms } => format!(
"no data for instrument {instrument} in window [{}, {}] (epoch-ms)",
window_ms.0, window_ms.1
),
MemberFault::Bind(detail) => format!("a member failed to bind: {detail}"),
MemberFault::Run(detail) => format!("a member failed to run: {detail}"),
MemberFault::Panic(detail) => format!("a member panicked: {detail}"),
}
}
/// Best-effort panic-payload message (String / &str payloads; else a fixed
/// marker). `AssertUnwindSafe` at the call sites is a documented judgement: the
/// runner is reused for later cells; the CLI runner holds read-only state.
fn panic_message(payload: Box<dyn std::any::Any + Send>) -> String {
payload
.downcast_ref::<String>()
.cloned()
.or_else(|| payload.downcast_ref::<&str>().map(|s| s.to_string()))
.unwrap_or_else(|| "member panicked (non-string payload)".to_string())
}
/// Assemble the (CellOutcome, CellRealization) pair for a cell that failed at a
/// stage (#272): the realized stage prefix survives in `stages`, `families`/
/// `selections` carry whatever already-persisted stages produced (empty for a
/// sweep-first fault; the sweep family/selection for a later wf-stage fault),
/// no nominee, the fault stamped on both, coverage from the runner.
fn failed_cell(
cell: &CellSpec,
stages: Vec<StageRealization>,
families: Vec<StageFamily>,
selections: Vec<StageSelectionOut>,
fault: aura_registry::CellFault,
runner: &dyn MemberRunner,
) -> (CellOutcome, CellRealization) {
(
CellOutcome { families, selections, nominee: None },
CellRealization {
strategy: cell.strategy_id.clone(),
instrument: cell.instrument.clone(),
window_ms: cell.window_ms,
stages,
regime: cell.regime,
regime_ordinal: cell.regime_ordinal,
fault: Some(fault),
coverage: runner.window_coverage(cell),
},
)
}
/// Sentinel `RunManifest::broker` value stamped onto [`placeholder_report`]'s
/// output — the single source both the setter and the survivor filter
/// (`run_members`'s `... != FAULTED_MEMBER_PLACEHOLDER_BROKER` sibling) read,
/// so the two can never drift apart into silently pooling a faulted
/// placeholder into a real family/OOS selection.
const FAULTED_MEMBER_PLACEHOLDER_BROKER: &str = "faulted-member-placeholder";
/// Slot filler for a faulted member: the engine sweep contract needs one
/// report per slot, but a captured fault aborts `execute` before any family
/// write, so this report is never persisted and never ranked.
fn placeholder_report(params: &[(String, Scalar)], window_ms: (i64, i64)) -> RunReport {
RunReport {
manifest: RunManifest {
commit: String::new(),
params: params.to_vec(),
defaults: Vec::new(),
window: (Timestamp(window_ms.0), Timestamp(window_ms.1)),
seed: 0,
broker: FAULTED_MEMBER_PLACEHOLDER_BROKER.to_string(),
selection: None,
instrument: None,
topology_hash: None,
project: None,
},
metrics: RunMetrics { total_pips: 0.0, max_drawdown: 0.0, bias_sign_flips: 0, r: None },
}
}
/// The shared in-sample-winner dispatch: called both for the sweep stage's
/// own family and for each walk-forward window's in-sample family. deflate=
/// true composes only with argmax (preflighted: `DeflatePlateauConflict`),
/// seeding the deflation nulls from the CAMPAIGN seed (C1: a campaign's
/// realization is a pure function of the document). Bare argmax has no
/// annotation to compute, so its `FamilySelection` is synthesized annotation-
/// free; `raw_winner_metric` mirrors the registry's ranking read (an R metric
/// against `r: None` ranks `NEG_INFINITY`). The plateau arms (`PlateauMean`/
/// `PlateauWorst`) ignore `deflate` outright (the `_` guard): they never
/// deflate regardless of the caller's flag, so the walk-forward caller's
/// hardcoded `deflate=true` only ever takes effect on the `Argmax` arm.
fn select_sweep_winner(
family: &SweepFamily,
axis_lens: &[usize],
metric: &str,
select: SelectRule,
deflate: bool,
seed: u64,
) -> Result<(SweepPoint, FamilySelection), ExecFault> {
match (select, deflate) {
(SelectRule::Argmax, true) => {
optimize_deflated(family, metric, DEFLATION_N_RESAMPLES, DEFLATION_BLOCK_LEN, seed)
.map_err(ExecFault::Registry)
}
(SelectRule::Argmax, false) => {
let winner = optimize(family, metric).map_err(ExecFault::Registry)?;
let raw = member_metric(&winner.report, metric).unwrap_or(f64::NEG_INFINITY);
let selection = FamilySelection {
selection_metric: metric.to_string(),
n_trials: family.points.len(),
raw_winner_metric: raw,
mode: SelectionMode::Argmax,
deflated_score: None,
overfit_probability: None,
n_resamples: None,
block_len: None,
seed: None,
neighbourhood_score: None,
n_neighbours: None,
};
Ok((winner, selection))
}
(SelectRule::PlateauMean, _) => {
optimize_plateau(family, axis_lens, metric, PlateauMode::Mean)
.map_err(ExecFault::Registry)
}
(SelectRule::PlateauWorst, _) => {
optimize_plateau(family, axis_lens, metric, PlateauMode::Worst)
.map_err(ExecFault::Registry)
}
}
}
/// A family point's tag-free cells lifted back to self-describing scalars in
/// `space` slot order — the survivor-point form the walk-forward stage
/// re-sweeps through a `ListSpace`.
fn scalar_point(space: &[ParamSpec], point: &[Cell]) -> Vec<Scalar> {
space.iter().zip(point).map(|(ps, c)| Scalar::from_cell(ps.kind, *c)).collect()
}
/// The pooled walk-forward OOS trade-R series: the family reports' `net_trade_rs`
/// concatenated in report order — which IS roll order (the wf stage builds its
/// family via `walkforward_member_reports_of`, per-window OOS reports in roll
/// order, over the surviving folds). Reports without an R block contribute
/// nothing.
fn pooled_net_trade_rs(reports: &[RunReport]) -> Vec<f64> {
let mut pooled = Vec::new();
for report in reports {
if let Some(r) = &report.metrics.r {
pooled.extend_from_slice(&r.net_trade_rs);
}
}
pooled
}
/// A placeholder for a faulted window slot: arity-correct chosen params (the
/// engine asserts every window's chosen-point arity against the param-space),
/// empty OOS equity, inert report. Never surfaces — a recorded window fault
/// fails the stage after the roll, before any registry write.
fn placeholder_window_run(specs: &[ParamSpec]) -> WindowRun {
WindowRun {
chosen_params: vec![Cell::from_i64(0); specs.len()],
oos_equity: vec![],
oos_report: placeholder_report(&[], (0, 0)),
}
}
/// Execute one `std::walk_forward` stage over the cell's surviving points:
/// roll (IS, OOS) window splits over the cell window — entirely in ms
/// (`Timestamp` is unit-agnostic `i64`; the driver owns ms->ns at its data
/// seam) — then per window sweep the survivor points over the IS bounds and
/// pick the winner via `select_sweep_winner` (the same dispatch the sweep
/// stage uses), seeded from the campaign seed: `Argmax` keeps the shipped
/// trials-deflation provenance, and `PlateauMean`/`PlateauWorst` read
/// `axis_lens` to score the closed neighbourhood — sound only because
/// preflight refuses plateau whenever a gate precedes this stage (a gate
/// filters survivors below the full grid, breaking the lattice `axis_lens`
/// assumes intact). Then run the winner over the OOS bounds with the
/// selection stamped on its fresh report, and append the per-window OOS
/// reports as a `FamilyKind::WalkForward` family.
#[allow(clippy::too_many_arguments)]
fn run_walk_forward_stage(
seed: u64,
cell: &CellSpec,
stage: usize,
block: &StageBlock,
specs: &[ParamSpec],
axis_lens: &[usize],
survivors: &[Vec<Scalar>],
family_name: &str,
runner: &dyn MemberRunner,
registry: &Registry,
) -> Result<WfStageOutcome, ExecFault> {
let StageBlock::WalkForward { in_sample_ms, out_of_sample_ms, step_ms, mode, metric, select } =
block
else {
return Err(ExecFault::PipelineShape {
detail: format!("stage {stage} is not std::walk_forward"),
});
};
if survivors.is_empty() {
// execute() truncates a cell at an empty gate before reaching here;
// refuse (never panic) if that contract is ever violated.
return Err(ExecFault::PipelineShape {
detail: format!("stage {stage}: walk_forward reached with zero survivor points"),
});
}
// Preflight guarantees the u64 lengths fit i64.
let span = (Timestamp(cell.window_ms.0), Timestamp(cell.window_ms.1));
let roll_mode = match mode {
WfMode::Rolling => RollMode::Rolling,
WfMode::Anchored => RollMode::Anchored,
};
let roller = WindowRoller::new(
span,
*in_sample_ms as i64,
*out_of_sample_ms as i64,
*step_ms as i64,
roll_mode,
)
.map_err(|e| ExecFault::Window {
stage,
// The detail string reaches user-facing seams verbatim — phrase both
// roller refusals in the doc's ms unit, never a Debug form.
detail: match e {
WalkForwardError::SpanTooShort { .. } => format!(
"in_sample_ms + out_of_sample_ms = {} ms exceeds the campaign window [{}, {}] ({} ms)",
in_sample_ms + out_of_sample_ms,
cell.window_ms.0,
cell.window_ms.1,
cell.window_ms.1 - cell.window_ms.0 + 1,
),
// Zero lengths are doc-tier faults (preflight); phrased anyway.
WalkForwardError::NonPositiveLength { field, value } => {
format!("walk_forward {field} must be > 0 (got {value})")
}
},
})?;
// A second identical roller enumerates the bounds up front: the engine's
// walk_forward consumes its roller, and the parallel closure needs each
// window's roll index for deterministic fault attribution.
let bounds: Vec<WindowBounds> = WindowRoller::new(
span,
*in_sample_ms as i64,
*out_of_sample_ms as i64,
*step_ms as i64,
roll_mode,
)
.expect("identical roller config validated above")
.collect();
let is_space = ListSpace::new(specs, survivors.to_vec()).map_err(|e| {
ExecFault::PipelineShape {
detail: format!("stage {stage}: survivor points do not fit the param space: {e:?}"),
}
})?;
let is_points = is_space.points();
// Windows run in parallel and the engine closure cannot return Err:
// faulted windows record here (keyed by roll index) and yield a
// placeholder run; after the roll the lowest-index fault wins
// (deterministic prose, the sweep-stage capture pattern one level up).
let faults: Mutex<Vec<(usize, ExecFault)>> = Mutex::new(Vec::new());
let result = walk_forward(roller, specs.to_vec(), |w| {
let widx =
bounds.iter().position(|b| *b == w).expect("bounds come from an identical roller");
// IS: engine sweep of the survivor points over the in-sample bounds,
// with the same member fault capture the sweep stage uses.
let is_faults: Mutex<Vec<(usize, MemberFault)>> = Mutex::new(Vec::new());
let is_family = sweep(&is_space, |cells| {
let params = zip_params(specs, cells);
let called = {
let _silence = SilencedPanic::enter();
std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
runner.run_member(cell, &params, (w.is.0 .0, w.is.1 .0))
}))
};
match called {
Ok(Ok(report)) => report,
Ok(Err(fault)) => {
let midx = is_points
.iter()
.position(|p| p.as_slice() == cells)
.expect("sweep enumerates exactly is_points");
is_faults.lock().unwrap().push((midx, fault));
placeholder_report(&[], (0, 0))
}
Err(payload) => {
let midx = is_points
.iter()
.position(|p| p.as_slice() == cells)
.expect("sweep enumerates exactly is_points");
is_faults.lock().unwrap().push((midx, MemberFault::Panic(panic_message(payload))));
placeholder_report(&[], (0, 0))
}
}
});
let member_faults = is_faults.into_inner().expect("no thread panicked holding the lock");
if let Some((_, fault)) = member_faults.into_iter().min_by_key(|(i, _)| *i) {
faults.lock().unwrap().push((widx, ExecFault::Member(fault)));
return placeholder_window_run(specs);
}
// IS winner: the same select_sweep_winner dispatch the sweep stage
// uses, seeded from the campaign seed. deflate=true keeps Argmax's
// shipped trials-deflation provenance identical to before this
// dispatch existed; the Plateau arms ignore the deflate flag.
let (winner, selection) =
match select_sweep_winner(&is_family, axis_lens, metric, *select, true, seed) {
Ok(picked) => picked,
Err(e) => {
faults.lock().unwrap().push((widx, e));
return placeholder_window_run(specs);
}
};
// OOS: run the winner over the out-of-sample bounds; the selection is
// stamped on the fresh OOS report (a new record, never a mutation of
// a stored one).
let oos_params = zip_params(specs, &winner.params);
let oos_called = {
let _silence = SilencedPanic::enter();
std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
runner.run_member(cell, &oos_params, (w.oos.0 .0, w.oos.1 .0))
}))
};
let mut oos_report = match oos_called {
Ok(Ok(report)) => report,
Ok(Err(fault)) => {
faults.lock().unwrap().push((widx, ExecFault::Member(fault)));
return placeholder_window_run(specs);
}
Err(payload) => {
faults
.lock()
.unwrap()
.push((widx, ExecFault::Member(MemberFault::Panic(panic_message(payload)))));
return placeholder_window_run(specs);
}
};
oos_report.manifest.selection = Some(selection);
WindowRun { chosen_params: winner.params, oos_equity: vec![], oos_report }
});
// Registry faults (optimizer/persistence broken) stay a global abort;
// Member/Window faults become per-fold WindowFaults via the same
// `contain` mapping the cell-level containment uses. One drain, in
// ascending window-index order — ExecFault/RegistryError are not Clone,
// so this is a single ownership-respecting pass rather than iter()-then-
// into_iter(); the first Registry fault encountered (ascending order =
// lowest index) wins, matching the existing lowest-index-fault convention.
let mut captured = faults.into_inner().expect("no thread panicked holding the lock");
captured.sort_by_key(|(idx, _)| *idx);
let mut window_faults: Vec<aura_registry::WindowFault> = Vec::with_capacity(captured.len());
for (widx, fault) in captured {
match contain(fault, stage) {
Ok(cf) => window_faults.push(aura_registry::WindowFault {
window_ordinal: widx,
kind: cf.kind,
detail: cf.detail,
}),
Err(global) => return Err(global),
}
}
// Keep only real OOS reports (drop faulted-fold placeholders by the broker
// sentinel) before building the family.
let surviving: Vec<&aura_engine::WindowOutcome> = result
.windows
.iter()
.filter(|w| w.run.oos_report.manifest.broker != FAULTED_MEMBER_PLACEHOLDER_BROKER)
.collect();
if surviving.is_empty() {
// Every fold failed — the cell fails at this stage.
let first = window_faults.first().cloned();
return Ok(WfStageOutcome::CellFailed(aura_registry::CellFault {
stage,
kind: first.as_ref().map(|w| w.kind).unwrap_or(aura_registry::CellFaultKind::Run),
detail: format!(
"all {} walk_forward folds failed{}",
result.windows.len(),
first.map(|w| format!("; first: {}", w.detail)).unwrap_or_default()
),
}));
}
let reports = walkforward_member_reports_of(&surviving);
let family_id = registry
.append_family(family_name, FamilyKind::WalkForward, &reports)
.map_err(ExecFault::Registry)?;
Ok(WfStageOutcome::Ran {
family: StageFamily { stage, block: "std::walk_forward", family_id, reports },
window_faults,
})
}
/// #272: a wf stage either ran (its family + any per-fold holes) or the whole
/// cell failed (every fold faulted). A Registry fault is a separate global Err
/// (returned directly from `run_walk_forward_stage`, never wrapped here).
enum WfStageOutcome {
Ran { family: StageFamily, window_faults: Vec<aura_registry::WindowFault> },
CellFailed(aura_registry::CellFault),
}
/// The surviving (non-faulted) walk-forward windows' OOS reports, in roll
/// order — the same mapping `walkforward_member_reports` does over a full
/// result, applied to the filtered survivor slice.
fn walkforward_member_reports_of(windows: &[&aura_engine::WindowOutcome]) -> Vec<RunReport> {
windows.iter().map(|w| w.run.oos_report.clone()).collect()
}