Semantic discovery audit
Inspected 2026-10-21 before this pass: README, AGENTS, Material navigation and technical/narrative pages; frozen contracts and generated schemas; discovery, fixture/MESA transport, Sentinel STAC planning and legacy acquisition; provenance, streaming gates, MOSAIKS, tests, examples, fixtures, tracks and decision history. The working tree contains earlier uncommitted review/MOSAIKS/website work; retain it.
Existing boundaries and concrete gap
discovery.run accepts a request and metadata-search client. It filters space,
time, modality and declared variable identifiers, returns candidate sources, and
does not approve equivalence. SourceResource already has identity, access, version,
rights, support and nested Metadata/Variable contracts. Variables carry definition,
unit, method, entity and one optional ontology ID. Native Sentinel metadata survives
as JSON with checksum scope, but semantic annotations have no typed extraction or
candidate-resource interface. MESA is explicitly unavailable; no API is guessed.
stream_sources.plan_sentinel reads STAC metadata only and proposes a fixed target;
calibration executes later in harmonization. Existing review/human gates must remain.
real_data snapshot acquisition is historical. provenance.export serializes full
source contracts; additive source semantics can therefore survive to representations.
Module 2 owns SemanticMapping; Module 3 consumes harmonized outputs only.
Focused response
Add optional nested source semantics with serialization preserving old identities when absent. Deterministically extract only declared metadata, distinguishing stored DN from documented calibration-output units. Return candidates without mapping or approval fields. Provide a local semantic-resource index and adapter protocol rather than downloading ontology graphs or binding core code to one portal.
Add a source-maintained landscape inventory, separately auditable project assessments, generated master table/matrix, glossary and review workflow. Retain all requested resource categories, including ambiguous historical leads with explicit unresolved identity/status/rights. Neither coverage scores nor lexical overlap proves scientific compatibility. The requested Parquet benchmark is a metadata diagnostic unless a separate scientifically reviewed evaluation protocol is supplied; no raw data reads.
Verification baseline
122 existing offline tests and a strict Material build passed before this pass. Existing fixture source checksums and old serialized plan/context identities need backward-compatibility tests. New tests cover partial/invalid contracts, annotation preservation, candidate provenance, Sentinel extraction and the discovery/mapping boundary. Full offline tests, generated reference, strict build and rendered links remain final requirements. No linter/type-checker is configured.