Runtime learning storage handoff

This completes the native learning, feature, hypothesis and control repository conversion following 0b90377. It also converts forward-evaluation persistence and the remaining direct learning-event reads. It does not select DynamoDB or S3 coordination authority or enable a hosted runtime.

Shared algorithms and native repositories

Domain	Public contract	Native implementation	Consumers
Learning artifacts and control-cycle progress	`LearningStorage`	`SQLiteLearningStorage`	Classifier fitting, question/example measurement, staged optimization and control scheduler
Feature hypotheses, lineage, attempts and deployment view	`FeatureStorage`	`SQLiteFeatureStorage`	Existing `FeatureBank` validation/algorithms
Retained control ideas	`ControlStorage`	`SQLiteControlStorage`	Control scheduler and shared hypothesis lookup
Hypothesis/diagnosis evidence records	`HypothesisStorage`	`SQLiteHypothesisStorage`	Existing `HypothesisLedger` statistics, comparison, lineage and diagnosis algorithms
Forward-evaluation membership, scoring and looks	`ShadowStorage`	`SQLiteShadowStorage`	Existing `ShadowEvaluator` admission, alpha spending and promotion policy
Runtime evidence/cycle reads	`RuntimeStateStorage.events_of` and `next_cycle_number`	`SQLiteRuntimeStateStorage`	Cycle replay, operational cycles and diagnosis

The core composes these repositories on its existing SQLite connection. Learning, feature, hypothesis, control, diagnosis, screening, proposal, cycle and forward- evaluation algorithms no longer query the wheel connection. wheel.db remains for native composition, existing embedded snapshots and local compatibility. There is no alternate partial-cloud constructor or second classifier algorithm.

The learning interface accepts five fixed artifact kinds (classifier, questions, examples, rubric, controls), never caller-supplied SQL table names. It retains raw progress/results, explicit retry behavior, newest-first artifact discovery and counts. Completion, promotion aliases and clearing an older pending-evaluation marker commit together. A failed alias insert rolls back all three changes. Existing answer-dependency validation remains in the shared engine.

FeatureBank(db) and HypothesisLedger(db) remain valid local constructors; storage= accepts the corresponding domain repository. Feature validation, lineage and diagnostics remain in FeatureBank. Statistical calculations, held-out exclusions, contextual evidence selection, verdict interpretation and operator/cause credit assignment remain in HypothesisLedger. Native adapters retain evaluation timestamp/row-order precedence and append-only verdict order. Absence of the optional control-ideas table still returns no idea; unexpected storage failures now propagate rather than being mistaken for missing evidence.

Forward-scored identities, joined order, old incumbents, statuses, and unique look identities remain native records. This extraction changes no screening partitions, review eligibility, request budgets, alpha spending or promotion rules. A cloud implementation must preserve the actual execution owner and existing uncertain-request rules; native transaction wrappers are not a distributed claim.

Complete per-classifier runtime migration

runtime_snapshot.py implements cyclotron-runtime/v1 for all 20 known native runtime tables: state/events, answers and retained history, learning trials, five learning-artifact kinds, features, control ideas, hypothesis evaluations and verdicts, diagnosis causes/elements and four forward-evaluation tables.

# Source is a quiescent native runtime store, or wheel.runtime_storage.
snapshot = source.export_storage(max_records=10000, max_bytes=64000000)

from decision_flywheel.runtime_storage import SQLiteRuntimeStateStorage
restored = SQLiteRuntimeStateStorage(destination_path)
restored.restore_storage(snapshot, max_records=10000, max_bytes=64000000)
restored.close()
# Open DecisionFlywheel with the matching task/model contract afterwards.

Lower-level export_runtime(db, ...) and restore_runtime(db, snapshot, ...) accept a native connection with no open transaction. Snapshot tables have fixed columns plus _rowid; present_tables retains lazy-table absence. Raw JSON stays text, preserving whitespace, Unicode, large integers, explanations and hashes. Physical row ordinals preserve equal-time evaluation precedence, append-only verdict order and newest-first artifact reuse. Runtime event and answer-history integer IDs preserve their rowids; these tables do not use AUTOINCREMENT and have no independent sequence watermark.

Export uses one SQLite snapshot and explicit record/byte preflights. Import first copies/validates the bounded snapshot, column types, JSON, ordinals, table presence and every native unique/key constraint in a trusted scratch schema. It never executes imported SQL. Unknown or changed table definitions, extra unique constraints, views and triggers fail closed. Restore requires an empty destination with compatible table presence, and creates/inserts the entire composition in one transaction. A mid-restore exception rolls back both new tables and rows. Restore is not a live merge or an authorization grant. Pending answers/trials remain pending; import does not replay or clear them.

RuntimeArchive uses the same CatalogArchive immutable per-record/manifest transport already used by console/catalog migrations, with a runtime-specific codec and table-presence metadata. File and injected-S3 round trips pass. Returned references are fixed to the authorized account/workspace scope; manifests are migration transport, not mutable runtime heads or DynamoDB authority.

This snapshot covers one classifier runtime database. It does not include the embedded Cyclotron root database, shared-batch database, console catalog/runs, reviewer source database, frozen-evaluation database, JSONL sidecars or sessions. Those require matching snapshots under host quiescence. A runtime export alone must never be advertised as a complete Cyclotron/workspace cloud migration.

Verification and remaining boundary

Focused consumer/repository regression: 199 tests passed. The final migration specifications passed 14 tests in a separate final run after adding injected-S3 transport and actual cache-only reopen coverage. These include complete 20-table round trips, raw JSON, row-order/timestamp ties, lazy schemas, interrupted restore rollback, concurrent multi-domain export consistency, bad schemas/identities/budgets, pending work and provider-free restart. A control/training integration test replaces wheel.db with a rejecting object and completes the three-control experiment using only the repositories.

An additional 100 native/embedded consumer tests passed. Final repository/ diagnosis/firewall checks passed 34 tests, including the first-human-feedback boundary with direct wheel SQL forbidden.

The integration owner's read-only execution harness passed 83 tests against this core and Papyrus source, including the public post-lock guard. No full suite, heavy build, cloud activation, IAM/OAuth change, public push or deployment was run.

The next independent conversions are embedded root operations/derived cursors, reviewer source state and frozen-evaluation clock/trace composition. Live cloud domain implementations still await the authority decision. Distributed ownership, worker lifecycle and session durability remain separate implementation/release requirements. Native history is intentionally not given a new silent cloud cap; cloud read/retention budgets must fail explicitly when selected.