# ADR-0129: Conserve represented interception supply

Date: 2026-11-01

## Status and provenance

Accepted for canonical integration and bounded final-engine/application
compatibility on 2026-11-02. The historical Roads-event hold below is superseded
by source-owned attribution, independent reviews and parent decisions. Optional
integration is separately accepted; repository validation and final publication gates remain open;
no deployment, retention default or universal physical closure is admitted.
User/Codex workspace conversation,2026-11-01,America/Los_Angeles. The user executed
the rock/water development program with authority for justified native ablations.
Decision owner: user for scope; scientific/implementation admission remains subject
to independent review. Implementer: Codex. No external meeting or endorsement implied.

## Observed problem

Unchanged-native A01/A02 replay reproduces a residue-limited IDAT allocation at
84543461 candidateKS1, year9/day355/OFE3. A rounded sum of melt plus carried canopy
water exceeds the represented primitive supply. Subtracting the proposed canopy
store transfers that excess to residue. The downstream signed budget is
-5.820766091346741e-11m before routing and fails its legitimate storage guard.
Evaluating that already overdrawn allocation in higher precision alone does not
repair its ownership. Actual source branch, zero irrigation and ordinary/full-
frost/legacy-observation parity are retained in the active program artifacts.

## Candidate decision

The initial A03 candidate used a DOUBLE sum and a downward REAL cast. It fixes
the observed history but fails constructed mixed-exponent and cancellation controls;
reject that algorithm for general admission. A04 instead carries the primitive
REAL supplies into an error-free binary64 expansion before comparing capacities
or computing remaining room. Preserve canopy-first/residue-second branch priority.
Certify accepted stores against the exact represented budget and certify maximal
REAL acceptance using its successor. Any canopy-limited representational remainder
continues as throughfall; it is not reassigned to residue. This uses no empirical
epsilon or minimum storage floor.

Strict operation ordering is required for the original expansion helper. Compiler
reassociation/unsafe math is disabled for its object. The finite binary32 primitive
range fits the binary64 expansion without overflow or underflow. Exact internal
allocation does not make every downstream REAL flux exact: returned binary64 and
persisted REAL remainders retain their explicit rounding boundaries.

Reconstruct signed remaining supply using the same budget-minus-accepted-stores
order. Preserve every representable residual as throughfall; do not discard it or
spend dry-reference soil water. IDAT's net interception and storm depletion,
both daily/hourly water-balance net-input calculations and ROCHEK's interception
subtraction must be assessed together. Preserve ROCHEK's effective-plane p and
geometry semantics; do not replace them with local rainfall/melt primitives.
Furrow irrigation remains excluded from interception and added at its existing
consumer. Signed downstream requests and their unmet-withdrawal guard remain.

## Alternatives and limits

Reject a FIN floor, arbitrary tolerance or removal of the storage guard. A residue-
only expression in current-input coordinates fixes the observed example but does
not resolve canopy-limited round-up or re-associated consumer subtraction. Global
REAL-kind changes introduce unrelated numerical effects. Limiting changes to a
shared original scalar allocation/remainder helper and confirmed consumers makes
the behavior testable without changing storage formats or scientific coefficients.

Standalone controls are necessary but insufficient. Require actual native replay,
complete calendars, producer/consumer readback, wider failure-class controls and
non-target regressions. Negative relative coordinates elsewhere and frost-policy
compatibility remain separate scientific questions. No general retention, surface
armor, frozen-water or physical-closure admission follows automatically.

## Evidence, compatibility and rollback

Active package: docs/work-packages/20261101_rock_water_development. A03 standalone
A03 controls compare2,148 bounded nonnegative-input cases against exact input fractions;
the old expressions generate negative local remainder in236 selected controls.
Those counts describe constructed controls, not field occurrence rates. Native
candidate changes remain private until reviewed integration/release decisions.
Default binaries, existing projects, stored soil files and user controls are
unchanged. Reject or revise a candidate on failed conservation, consumer/domain
or regression gates; retain every failed lane and restore the accepted comparator.

## Execution evidence and remaining admission gates

A03-v2 completes the formerly failing16year history with17,532 daily records and
an exact zero target remainder. A04-scalar-v3 passes20,839 rational-oracle cases,
including both numerical counterexamples, subnormal-to-maximum finite exponent
coverage, signed negative consumer controls and actual failing primitive operands.
This is finite coverage, not exhaustive all-tuples enumeration.

A04-v2 completes all twelve retention/KS cases. Zero-rock baseline/candidate outputs
are exact under both flag settings. Neither result admits general physical water
closure. The canopy carry-transfer branch may reassign already allocated carry;
its actual execution/ownership must be observed before repair. Storm depletion,
PERC domains, accepted input/percolation, full-project and native regression gates
also remain. The original DOUBLE-only candidate remains a rejected general method;
the expansion candidate is under investigation and has not been released.

## Integration admission, 2026-11-01

Independent N01 scientific and implementation reviews admit the exact A04 helper
and four shared consumers, with strict compilation and complete regression tools.
A06 completes 64,935,267 observed domains without an invalid domain; A07 confirms
selected storm ownership without demonstrating a runoff-producing carry defect.
The isolated general candidate passes 1,260 native tests, both host smokes and
31 watchlist cases. Two fresh nine-year medium watersheds retain all inputs and
output filenames; outlet volume changes by -0.82 m3 and sediment by -2.41 kg.
These are bounded regression observations, not general physical-closure proof.

The exact reviewed source/test/tool/fixture set is integrated on the canonical
native branch; no binary is replaced yet. Earlier candidate-state descriptions
above record the investigation chronology. The user authorizes re-releasing the
undeployed Linux wepp_261029 matched pair and publishing revision-5 notes locally
in WEPPpy after remaining Roads, Disturbed and exact-release gates. Retention
defaults, frost policy, deployment and Windows publication remain outside this
admission. Any adoption must regenerate the matched hillslope/PASS/watershed chain.

Review records: active program artifacts/N01-candidate-science-admission.md and
artifacts/N01-regression-v2-implementation-outcome.md. The integration receipt is
retained privately under N01-integration-v1/source-manifest.json.

## Release hold after full Roads readback

The final pair passes 1,296 native tests, both host smokes and the 31-case watchlist.
All 840 Roads histories complete, with the original magnitude classification
unchanged. Those results do not establish physical admission: one local event
changes runoff by +39.003874 mm and liquid storage by -38.94 mm; separate cases
change lateral flow by +17.65 mm and uptake by +4.05 mm. Two daily visible-water
ledgers exceed their printed-value uncertainty, but omit stores or stage ownership
needed to diagnose water creation or loss.

Keep the release on hold until paired source observations explain these changes
and independent scientific and implementation reviews resolve their disposition.
Final reported upstream peak/duration can be overwritten during downstream OFE
processing; it cannot alone identify the forcing that caused the difference.
Preserve the admitted allocation invariant while investigating the complete
producer/consumer chain. Any additional repair requires its own evidence and
renewed affected gates; do not weaken thresholds to admit the current pair.

Evidence: active program artifacts/R02-science-readback.md,
R05-water-extremes-triage.md and R05-hydrograph-source-scope.md. The authorized
destination remains the undeployed wepp_261029 revision 5 in WEPPpy. No replacement
binary, release report or deployment has occurred at this checkpoint.

## Current bounded admission, 2026-11-02

The exact expansion allocation and confirmed IDAT, daily/hourly FIN and ROCHEK
consumers are admitted as a bounded numerical water-ownership repair. A05 records
30,974,527 executed domain observations with no flagged domains; A06 records
64,935,267 with no flagged domains, joins actual producers to consumers, and
retains selected accepted-flux checks and negative relative-coordinate flags.
These counts are evaluations, not independent states or whole-history closure.
A07 observes changed storm/depletion versus later carry allocation, but its
selected event does not enter GRNA/ROCHEK or prove water creation. No additional
carry fallback or timing repair is admitted from that observation.

Final source `d2e14d96af5942409c9a315dc8594fcc164a702a` combines this repair with
the separately admitted snow-pack debit and donor-limited freeze-restoration
repairs (ADR-0130 and ADR-0131). Combined18, 1,394 native tests, 31 watchlist
cases, two host smokes, six legacy histories, Disturbed, clean and observed
Roads840 panels, and the 492-history application comparison have completed
bounded independent review and parent admission. The exact matched pair is
installed locally with rollback evidence; deployment and final publication
remain separate. These later decisions supersede the historical release hold,
not its retained observations.

Roads engine compatibility and capacity/particle arithmetic are accepted; the
original parameterization screen still fails at 744/840. No threshold was relaxed
and no field calibration follows. Application changes include 978 of 3,443
files, outlet volume −1,129.18 m3 and sediment −49.22 kg, with unchanged maximum
peak. Material storage/lateral timing, liquid/ice partition and channel-area
consequences remain disclosed. Neither these comparisons nor unchanged outputs
prove universal thermal validity, cold-snow budgets, whole-system conservation
or retention defaults. Optional retention has separate bounded P03/P04 and installed-controller
admission; its validity must not be inferred from this numerical repair.

Current parent decisions and independent evidence are retained in the active
program: `combined18-final-execution-v3/parent-acceptance-v1.json`,
`legacy-frost-final-renewal-v1/parent-acceptance-v1.json`,
`final-roads-parent-disposition-v1/decision.json` (SHA-256
`fb384407f3d7cf85046e24103ef5fa15e0b415e71d3c9ef5e18579ef4f18c473`), and
`final-application-parent-disposition-v1/decision.json` (SHA-256
`615cedebc5ad2c19e3ca18ac930030f4a9f7aef7decd1b2cdf0a50b59d298194`).
The latter binds actual science, implementation and security reviews. The
program's N01/A05/A06/A07, combined18, native, legacy, Disturbed, final Roads
and application review artifacts retain the precise tested scopes and limits.
