# ADR-0076: Continuous watershed runtime budget

Date: 2026-09-28

## Status and provenance

Accepted, 2026-09-28. Implemented in `aa2d1cd2a`. Decision venue: this
Codex/operator conversation (America/Los_Angeles). Participants: operator and
Codex. Decision owner: operator (“let's implement the year plus hillslopes timeout
please”). Implementer: Codex. Existing session commit authority applies.

## Decision

Replace the fixed 12-hour continuous watershed child allowance with 0.05 seconds
per hillslope-year, rounded up to whole hours, retaining 12 hours and any larger
caller pipeline allowance as floors. Use controller workload for preparation,
checked-out native run-file workload for no-preparation. Preserve all unrelated
stages, single-storm limits and graph dependencies. Record budget inputs in child
metadata. Exact contract: [WRT-01](../schemas/wepp-run-input-contract.md#continuous-watershed-runtime-budget-wrt-01).

## Rationale and alternatives

[Read-only wepp1 evidence](../investigations/20260928_wepp1_watershed_timeout_scaling/assessment.md)
matched 123 successes to consumed artifacts. Years plus hillslopes explained
runtime much better than years alone; channels were nearly collinear with hills.
The observed same-binary 95th-percentile rate was 0.047 seconds per hillslope-year.
A 1,908-hillslope 500-year run took 8.76 hours; 1,000 years reached year 698 at 12 hours.
The policy allows 27 hours for the latter. Reject global timeout increases and
separately fitted channel weights. No scientific inputs/formulas are changed.

## Compatibility, risk and rollback

Metadata is additive. Valid prepared no-prep inputs take precedence over stale
NoDb values; legacy and modern prompt layouts remain supported. Invalid workload
fails before enqueue. Longer jobs occupy workers longer; the measured sample
is mostly one binary and 100-year runs, with successful-job selection bias.
Runtime cleanup on timeout remains a known separate issue. No new dependencies,
services, automatic retries or deployment are authorized by this change.
Rollback restores enqueue defaults for new jobs; existing queued/failed jobs
retain their saved timeout. Do not mutate those jobs as part of rollout.

## WRT-02 Omni leaf scope amendment

Accepted 2026-10-06 in this Codex/operator conversation. Participants were the
operator and Codex; the operator requested that Omni scenarios and contrasts use
the dynamic timeout, and Codex is the implementer. This amendment applies the
unchanged WRT-01 options to newly submitted `run_omni_scenario_rq` and
`run_omni_contrast_rq` leaves in the existing job-pool concurrency path. It does
not change the coefficient, floor, rounding, metadata, or range. It is a
workflow-scope application/default amendment: the covered Omni leaves change
from fixed to workload-derived allowances using the accepted numeric policy.
Implemented locally in `adff42d6d9e38f561d41cc681e89f3386ef888c0`;
deployment remains separate.

The rejected alternatives were a global Omni timeout increase, which would not
track workload, and a new Omni overhead coefficient, for which no retained
empirical fit exists. Coordinator, compile, finalizer, retry, cleanup and
deployment behavior remain outside the amendment. If environment evidence shows
that total Omni leaf overhead exceeds the unchanged budget, retain that evidence
and authorize a separate parameterization decision rather than silently raising
this allowance.
