ADR-0143: Candidate nominal 1 m culvert pruning boundary
Status: Superseded candidate by ADR-0144
Date: 2026-11-16
Context
Santee's nominal 1 m DEM reports 1.0000000000089009 m, selecting two pruning passes and producing 14,339 subcatchments and 5,765 channels.
Decision
Evaluate a 1.5 m inclusive upper boundary for three passes on forest. Acceptance remains pending a fresh RQ replay of Point_ID 1040. Both counts must be within half to twice the retained 726 subcatchments and 298 channels, with valid connected drainage/outlet and no silent unpruned-stream fallback.
Decision Provenance
Decision Venue: Operator/Codex conversation, recorded 2026-11-16 00:46 UTC.
Participants Present: Operator and Codex.
Decision Owner(s): Operator.
Implementer(s): Codex.
Change Summary
Effective resolution remains cell size times sqrt(flow_accum_threshold / 100). Three passes change from effective resolution <= 1.0 m to <= 1.5 m; two passes cover > 1.5 through 4 m; one pass covers > 4 m. Missing or nonpositive thresholds still use 100. Explicit payload overrides remain intact.
Rationale
Represent the nominal resolution class explicitly instead of treating metadata precision as a scientific distinction. This is a heuristic pending calibration.
Alternatives Considered
- Epsilon only: repairs rounding but preserves an unsupported 1.01 m cliff.
- Floor: also reclassifies genuine 1.9 m and 4.9 m data.
Consequences
There is no retained intermediate 1.5-2.5 m production calibration site. The existing 2 m test is regression evidence only. Santee acceptance cannot establish general calibration for every resolution or watershed.
Evidence
See the execution plan for retained batches, acceptance gates, and replay procedure.
Risk and Rollback Notes
Three passes may still produce excessive counts or lose drainage semantics. Reject acceptance and revert to the strict mapping if replay gates fail. Do not promote to production pending forest evidence and final decision.
Implementation Notes
Preserve the helper signature, formula, 4 m boundary, and queue wiring. Require uncached SDA soil builds before restart and immediately before replay.
Disposition
The three-pass candidate repaired classification but failed the Santee count criterion. The operator subsequently selected per-run coarse-to-fine pruning in ADR-0144, preserving supplied culvert connections while preferring simpler networks. The earlier experiment remains evidence, not the final default policy.