Model materialization evidence

Issue #39 extends P07 with temporary stored inputs and child revision publication. Materialization means staging validated inputs for temporary use. The tested source is 32fd79e7238a7dad416da4e5be7318ca2febecf3. The seven changed files match the reviewed change ef3b6910c9b435046c6117dd8f11fd588a1aab44, including its cleanup fixes. The import guide describes the public interface and ownership transfer.

Public behavior

The public library tests use the stored model reference after deleting the original source directory. They verify exact revision identity, independent staging directories, and removal of staged inputs and the property file after exit. Changing one staged copy leaves another copy and the stored parent unchanged. Missing revisions and input files outside the stored include graph fail before the context yields files.

The isolated OPM parser supplies cell order, geometry, static properties, and report times. OPM is the Open Porous Media project. The tests verify I-fastest indices, an 8,335-foot cell depth, and a 20,000,000-cubic-foot cell volume. They check a 1,000-foot cell length, 500-millidarcy permeability, and report days zero, one, and two. They parse the derived property file and compare all seven arrays with the inspection. The arrays contain DX, DY, DZ, PORO, PERMX, PERMY, and PERMZ in FIELD units.

Child publication inherits the stored parent's session, coordinates, and units, and records its parent reference. The tests change a schedule value and verify that the child contains it while the parent keeps the original value. Invalid child inputs publish no artifacts. An injected child publication failure preserves the parent and reports UNKNOWN after artifact writes begin. This result requires recovery inspection and does not promise rollback.

Profile and failure checks

New validation uses spe1-field-v2. An existing spe1-field-v1 summary still deserializes without rewriting its profile. The pinned OPM version remains 2025.10, and supported black-oil physics stays unchanged. The tests compare explicit connection factors and permeability-length values with OPM schedule values in its standard units. Invalid factors, permeability-length values, skin values, directions, and completion ranges fail validation. Inactive-cell inputs remain outside the supported profile.

A consumer's OSError passes through as the same exception, and staged files are removed. If cleanup also fails, the same consumer exception retains its mutation effect and receives a cleanup note. If the consumer succeeds but cleanup fails, materialization reports STORAGE_FAILED with effect UNKNOWN. A regression test invokes the context inside an unrelated exception handler. It verifies that cleanup failure remains visible and does not attach a note to the unrelated exception.

Validation record

The environment record preserves the source commit, working state, versions, and commands. The shared check log records Ruff, ty, pytest, and the strict documentation build. The issue snapshot preserves the reviewed scope. The dependency log records the locked environment installation. The shared command passed all 427 tests, Ruff, ty, and the strict documentation build. The result record separates those outcomes from external acceptance. The test output establishes public library behavior through the real workspace store and isolated parser.

The integrated check passed 453 tests, Ruff, ty, and strict documentation after P08 merged. The integration record identifies its source base and working changes. The current guides name profile version 2 while preserving version 1 in historical P08 trial records.

The browser record checked two local pages and eight local links without page errors. The visual review records inspection of both captured screenshots. The headings, text, and materialization table remained readable without clipping or overlap. The integrated browser review covers the navigation, current profile, capability map, and linked evidence.

The review assessed duplicate behavior, control flow, ownership, failure handling, tests, and evidence limits. Original imports and child revisions share one validation and publication path. The materialization context owns staging cleanup and keeps consumer failures distinct from service failures. The changed public tests exercise observable values and failure outcomes. No blocking review finding remained.

Evidence limits

This local check does not replace Linux and macOS CI. It does not launch ResInsight, run a simulator, or exercise MCP wiring. MCP is the Model Context Protocol for tool access. Native input-grid and well acceptance belong to P09 under issue #10. Those separate checks establish whether native consumers use these staged files correctly. This record does not infer native success from library tests.