feat(orchestrate): EdgeDecayTrigger, the parameter-space half of monitoring

[mspinola]

Warning

Blocked on a crucible release. This raises the crucible floor to >=0.5.0, and unlike the 0.3.0 constraint that one is not yet true: the monitor sits in crucible's [Unreleased] and the newest published crucible is 0.4.0. Merging is fine; releasing crucible-stack from this before crucible 0.5.0 is on PyPI would make import crucible_stack.orchestrate raise for anyone installing from PyPI.

Why

drift watches the equity path: cumulative R and drawdown against the envelope frozen at promotion. Nothing watched the per-trade parameter. They are complements, and they come apart in both directions:

• a halved expectancy leaves the path flat, and flat usually sits inside a p5 band provisioned over a multi-year horizon
• a run of correlated losers breaches the drawdown floor with every per-trade statistic exactly where it should be

Running only one leaves a real failure mode invisible.

What

The judging is crucible's (crucible.validation.monitor). What this adds is the three things crucible refuses to own:

1. freezing an EdgeBaseline at the moment of promotion
2. persisting it in the DeploymentLedger beside the DriftEnvelope
3. turning a verdict into a re-optimization trigger

Same division of labour as drift, whose module header already reserved this shape ("designed to migrate into crucible if it earns its way").

Only DEGRADED fires. crucible reserves that label for the CUSUM, the one channel carrying a stated false-alarm rate. SLIPPING is reported in reasons and does not trigger. Letting it trigger would import an uncalibrated tripwire into the loop, and would tax the honest N besides, since every re-optimization is variants added to the SearchSpaceLog — the trigger docstring already makes that argument about spurious fires.

A refusal carries no baseline, matching the envelope rule, so there is never a reference a later cycle could re-baseline onto.

Three things the existing tests forced, all improvements

test_triggers_know_nothing_about_the_substrate pins trigger.py's import surface to crucible_stack + numpy. My first cut imported crucible directly and broke it. Rather than widen the allowlist to fit new code, trigger now depends on decay.check_decay, exactly as it already depends on drift.check_drift. The layering reads better for it, and the guard stays as strict as it was.

test_public_api pins the exported surface, so the new names are added there in the same commit rather than becoming an accidental promise.

test_crucible_compat is the repo's stated real enforcement ("a version number is not evidence that an API is present"). It gains a check for the monitor surface, plus one asserting edge_monitor still has no parameter from which a baseline could be rebuilt. That second one matters: a future crucible relaxing that guard would silently make EdgeDecayTrigger incapable of firing while every other test here kept passing.

A units trap worth naming

TriggerContext gains trade_r (per-trade R) as a field separate from realized_r (per-period R, the grid the envelope was built on). Two series, two lengths, two clocks (trades_live vs elapsed). Reusing one field for both would have been the units bug crucible fixed in v0.4.0, where an observed expectancy in R was compared against a null in fractional returns.

run_cycle's trade_r is optional and defaults to empty, so existing callers are unaffected.

Verification

• 388 tests pass (+23), against crucible main via PYTHONPATH
• ruff check crucible_stack tests: clean

Note that crucible-stack's local .venv has crucible 0.1.0 in site-packages — the unrelated 2011 PyPI squatter, not this project — so it cannot import the real crucible at all. That is a pre-existing local environment problem, not something this branch introduces, but it does mean local runs need PYTHONPATH=../crucible/src until the venv is repaired.

Not done

npf needs no change: npf/scripts/orchestrate_trend.sh already invokes python -m crucible_stack.orchestrate, so it picks this up for free. What npf's adapter would eventually supply is the trade_r series and a Reoptimization.baseline; until it does, EdgeDecayTrigger fails open with an explicit reason rather than pretending the book is healthy.

🤖 Generated with Claude Code