Feat/2748 paginated seed browsing

[Jackson Severino da Rocha (jacksonsr451)]

Description

Implements #2748 by adding typed, read-only APIs for browsing and retrieving logical seed examples from existing seed memory.

This builds on the dataset identity and count contract introduced in #2762 and follows the browsing semantics discussed with Roman Lutz (@romanlutz) in #2748.

What changed

• Added paginated logical seed example browsing with opaque, filter-bound keyset cursors.
• Added deterministic ordering by the earliest stored date_added of the complete logical example, descending, with logical-example ID as the tie-breaker.
• Uses prompt_group_id as the logical-example identity, falling back to the persisted seed ID for ungrouped records.
• Added database-bounded filtering for:
  • modality
  • harm category
  • seed type
  • stored text search
• Applies OR semantics within each filter family and AND semantics between filter families.
• Evaluates filters at the logical-example level while returning all members of a matching example.
• Added case-insensitive whole-value harm-category matching.
• Added literal case-insensitive text search with % and _ treated as literal characters.
• Added compact list projections with bounded previews and full detail projections for individual logical examples.
• Added logical-example totals using the same predicates as the paginated query.
• Added typed list/detail response models and HTTP routes:
  • GET /api/datasets/{selection_key}/seeds
  • GET /api/datasets/{selection_key}/seeds/{example_id}
• Preserves persisted seed IDs, nullable group IDs, roles, sequence values, hashes, provenance, metadata, objective fields, template parameters, and persisted template state.
• Added nullable persistence for is_jinja_template so browsing can distinguish templates without inferring template state from parameters.
• Added side-effect-free projections for browsing without reconstructing seeds or groups.
• Added direct bounded detail lookup rather than scanning paginated results.
• Added SQLite and Azure SQL implementations for the required browsing predicates.
• Preserved the existing MemoryInterface.get_seeds() behavior.

Safety and side-effect behavior

Browsing does not:

• render templates;
• reconstruct legacy simulated conversations;
• load referenced template files;
• load media contents;
• generate conversations;
• call models or targets;
• write to seed memory;
• fetch provider datasets;
• load provider metadata solely to validate browsing selections.

Named browsing selections are resolved from datasets already represented in seed memory. A provider-only dataset that has not been persisted to memory is therefore not treated as browseable by these endpoints; this avoids provider/file loading during a read-only browsing request.

List previews also avoid exposing absolute filesystem paths or credential-bearing URLs.

Fixes #2748

Tests and Documentation

Added and expanded contract coverage across the Memory, service, persistence, and HTTP layers.

Coverage includes:

• bounded database pagination;
• deterministic ordering;
• tied timestamps;
• cursor continuation;
• invalid and filter-mismatched cursors;
• complete grouped-example retrieval;
• cross-member filter matching;
• OR-within-filter / AND-between-filter semantics;
• whole-value harm-category matching;
• literal % and _ text searches;
• named and unnamed dataset isolation;
• unknown dataset handling;
• preview truncation;
• full-detail fidelity;
• original seed/group identifiers;
• nullable template state;
• template parameters;
• side-effect-free browsing;
• provider/file-loading guards;
• direct bounded detail lookup;
• SQLite and Azure SQL query behavior;
• migration behavior and historical nullable template values;
• HTTP dependency isolation.

Final focused validation:

• HTTP seed browsing contract: 40 passed
• Dataset service tests: 28 passed
• Memory browsing validation: 108 passed
• Migration tests: 71 passed
• Relevant backend regressions: 158 passed
• Ruff: passed
• git diff --check: passed

Azure SQL integration was also validated against a real Azure SQL database after the SQL/schema corrections:

• Azure SQL seed browsing integration: 1 passed

The final HTTP/service changes did not modify the SQL query, schema, or migration behavior after that Azure SQL validation.

JupyText was not run because this change does not modify notebooks or JupyText-managed documentation/examples.

[Jackson Severino da Rocha (jacksonsr451)]

@microsoft-github-policy-service agree