feat(evaluation): bridge installed datasets into production runner

[daniele21]

Scope

Adds the concrete production dataset-to-runner bridge required to move EVAL-2/EVAL-4 beyond fake case sources without coupling evaluation:engine to filesystem or parser implementations.

• introduces evaluation:dataset-adapter as a concrete integration module depending on evaluation:datasets and evaluation:engine;
• implements the existing runner EvaluationDatasetPreflight and EvaluationCaseDefinitionSource ports from registry-published installed packs;
• resolves only the exact dataset ID/version from EvaluationDatasetRegistry and verifies the requested content digest;
• reparses cases.jsonl through the canonical bounded EvaluationDatasetJsonlParser and verifies canonical content digest before exposing cases;
• validates sample-set dataset identity and rejects selected case IDs not present in the installed pack;
• caches only a validated immutable in-memory snapshot keyed by exact dataset identity, so scored cases do not reparse the file repeatedly;
• exposes manifest category definitions from the same validated snapshot for the subsequent R-10 aggregation wiring;
• adds deterministic tests for production pack consumption, exact case lookup, category exposure, digest mismatch, invalid sample membership and missing packs;
• registers explicit module ownership/navigation and keeps dependency direction one-way through the adapter.

Boundary

This PR does not add filesystem access to evaluation:engine, does not create an alternate parser, and does not make evaluation:datasets depend on the engine. The adapter is the composition seam between those two existing ownership boundaries.

This slice also does not yet wire EvaluationRunAggregator into the engine terminal path; the exposed category definitions are the input for that next composition slice after R-09 converges.

Gate

Keep draft until module registration/navigation, scoped adapter tests/lint and repository validation are green. EVAL-2 should not be marked DONE until this bridge is integrated and used by production evaluation composition.