Research Backlog (Adapted from January 2026 agenda)¶

This backlog turns the 20 research proposals in docs/RESEARCH_AGENDA.md into Codex-friendly, PR-sized tasks, mapped onto the existing Phase A–G build plan and the current docs/PROJECT_BACKLOG.md task numbering.

Notes: - These items are research-oriented: the goal is to make each direction implementable and measurable. - Tasks are designed to be optional and do not block core MySQL compatibility.

Reality sync (2026-05-27)¶

This file is the research task inventory. It is not the best place to read current maturity at a glance.

• Runtime truth: all R01-R20 tracks have executable coverage in code, methods, tests, or benchmark scaffolds.
• Current maturity split: R01-R17 and R20 are hardened; R18 and R19 remain prototype implemented.
• Checklist state: 109 done / 0 open research checkboxes. The checklist is complete, but some tracks still remain prototype-strength in runtime maturity.
• Current source of truth for implemented-vs-partial status: docs/TRUE_STATUS_MATRIX.md.

Current partial research areas¶

Recent verified closures¶

• 2026-05-17: R02 adaptive row/column hardening was finished across T103-T106, covering live-row-count cost modeling, hot-projection tracking, dependency-driven snapshot invalidation, and adaptive replacement decisions.
• 2026-05-17: R14 bounded-staleness bundle windows were closed with retained coverage-gap detection in edge.bundle.status.
• 2026-05-17: R16 workload-shift evaluation was closed with phased convergence reporting in advisor.evaluate.
• 2026-05-17: R18 deterministic replay reconstruction was closed with replay cache-hint rehydration and normalized replay-run checksums.

Use the task sections below for the detailed per-track inventory and evidence notes.

Mapping table (proposal → repo)¶

Task definitions (new additions)¶

Phase 24 — Learned indexes for ValueID lookup (R01)¶

• [x] T230: Instrument ValueID lookup distribution + export histograms
• [x] T231: Prototype learned model index (offline build) with fallback structure
• [x] T232: Integrate hybrid learned+fallback lookup into ValueStore read path (feature flag). Evidence: ValueStoreConfig.enable_learned_index, ValueStore::get_with_trace, and learned_index_lookup_hits / learned_index_falls_back_for_new_keys.
• [x] T233: Compaction-time model refresh policy + correctness tests. Evidence: ModelRefreshPolicy, ValueStore::should_refresh, maybe_refresh, refresh_learned_index, insert-triggered refresh checks, and distribution_shift_triggers_refresh.
• [x] T234: Benchmark harness: lookup p50/p99/p99.9 + memory overhead. Evidence: ValueStore::benchmark() and benchmark_reports_quantiles.
• [x] T235: Distribution shift tests + graceful degradation. Evidence: ValueIdLookupDistribution::model_shift_l1, learned_index_falls_back_for_new_keys, and distribution_shift_triggers_refresh.

Phase 25 — Differential privacy aggregates (R04)¶

• [x] T240: Add SkeinQL aggregate nodes (COUNT/SUM/AVG) with explicit DP parameters (experimental). Evidence: dp.aggregate, DpAggregateSpec, COUNT/SUM/AVG result columns, explicit epsilon/delta/mechanism/principal/seed fields, and dp_budget_consumption_and_exhaustion / dp_aggregate_deterministic_noise tests.
• [x] T241: Sensitivity analysis for single-table aggregates (bounded domains). Evidence: resolve_dp_aggregates, bounded DpBounds range sensitivities for SUM/AVG/percentile, count sensitivity 1.0, privacy metadata per aggregate, and focused assertions in dp_budget_consumption_and_exhaustion.
• [x] T242: Privacy budget manager (per user/role) + persistence. Evidence: dp.budget.set, dp.budget.get, DpBudgetDisk v2 persistence in dp_budgets.json, refresh-window resets, RDP query counts, and restart assertions in dp_budget_consumption_and_exhaustion.
• [x] T243: Noise mechanisms (Laplace / Gaussian policy) + deterministic tests (seeded RNG). Evidence: DpRng, dp_laplace_noise, dp_gaussian_noise, mechanism validation, seeded deterministic Laplace/Gaussian coverage, and r04_dp_rng_deterministic_and_uniform / r04_dp_laplace_noise_has_correct_scale / dp_aggregate_deterministic_noise tests.
• [x] T244: Privacy-aware caching rules (ETag includes privacy metadata). Evidence: privacy_etag in dp.aggregate privacy output, derived from a v1 DP validator payload containing table version, query fingerprint, epsilon/delta, mechanism, principal, seed, and budget metadata; locked by dp_budget_consumption_and_exhaustion.
• [x] T245: Audit log entries for DP queries (budget consumption). Evidence: DpAuditEvent, dp.audit.log, persisted dp_audit.json, budget remaining epsilon/delta in events, usage summaries in dp.budget.get, and restart assertions in dp_budget_consumption_and_exhaustion.
• [x] T246: Evaluation harness: accuracy vs epsilon, overhead vs baseline. Evidence: dp.evaluate, DpEvaluateParams / DpEvaluateResult, exact baseline rows, seeded epsilon-grid trials, mean/p95/max absolute error, mean relative error, noisy latency, overhead-vs-exact metrics, SkeinAdmin Privacy controls, and dp_evaluate_reports_accuracy_and_overhead / skeinadmin_privacy_panel_exposes_dp_evaluation_harness tests.

Phase 26 — Oblivious query execution (R05)¶

• [x] T250: Threat model doc + “obliviousness levels” policy schema. Evidence: docs/OBLIVIOUS_EXECUTION.md, ObliviousPolicy, oblivious.policy.set, normalize_oblivious_policy, and persisted ObliviousPolicyDisk v1.
• [x] T251: ValueStore lookup padding + dummy reads (table/column policy). Evidence: oblivious_padding_for, oblivious_dummy_lookups, compute_oblivious_padding, and oblivious_scan_keeps_results.
• [x] T252: Oblivious scan primitive (fixed-size batches, padding). Evidence: scan_table applies deterministic padding/shuffle before returning real rows unchanged, locked by oblivious_policy_explain_padding and oblivious_scan_keeps_results.
• [x] T253: Oblivious sort/join primitive (limited scope, research mode). Evidence: oblivious.explain / oblivious.evaluate report materialize_then_sort_join for padded policies and expose target/dummy access envelopes for fixed-size inputs.
• [x] T254: Leakage evaluation harness (trace-based, mutual information metrics). Evidence: oblivious.evaluate, ObliviousEvaluateResult, empirical mutual-information metrics, and engine/RPC assertions comparing padded vs unpadded traces.
• [x] T255: Performance overhead report generator. Evidence: oblivious.evaluate performance payload with mean/max overhead ratio, total dummy rows/lookups, total observed accesses, and integration coverage in r05_oblivious_padding_verification.
• [x] T256: Admin UI settings for per-table obliviousness levels. Evidence: SkeinAdmin Privacy R05 controls for level/pad/target/dummy/shuffle/trace rows, oblEvaluate(), and skeinadmin_privacy_panel_exposes_dp_evaluation_harness asset coverage.

Phase 27 — Forensic query language (R06)¶

• [x] T260: Define SkeinForensic query grammar (minimal) + JSON form over SkeinQL. Evidence: ForensicQueryParams.filter, forensic_filter_matches, operators and/or/not/eq/ne/gt/ge/lt/le/contains, typed-literal operands, field equality shorthand, docs in docs/AUDIT_WAL.md, and focused engine/RPC tests.
• [x] T261: Build verifiable WAL index (time/table/user) consistent with hash chain. Evidence: forensic_index_summary emits timestamp/id ranges, by_table, by_op, and by_actor buckets tied to the returned chain/proof; actor remains unknown until authenticated principal metadata is recorded.
• [x] T262: Proof format for inclusion + boundary proofs; verifier tool. Evidence: skein.forensic.proof.v1, boundary preceding_hash/following_hash, forensic_merkle_root, forensic_merkle_proof, per-record inclusion_proofs, and forensic.verify tamper detection.
• [x] T263: forensic.query SkeinQL endpoint + exportable report bundles. Evidence: JSON-RPC dispatch, capability advertising, skein.forensic.bundle.v1 query manifest/proof/verification export shape, and RPC roundtrip coverage.
• [x] T264: Incremental verification via checkpoint anchors. Evidence: persisted CheckpointAnchor records and proof fields checkpoint_anchor, next_checkpoint_anchor, and anchor_count, with engine coverage after checkpoint_for_shutdown().
• [x] T265: Case-study harness: simulated incident timelines + proofs. Evidence: forensic_case_study_exports_incident_timeline covers non-contiguous filtered incident timelines, inclusion proofs, and export-bundle verification strategy.
• [x] T266: SkeinAdmin “Forensics” page (query + verify + export). Evidence: DB/table/op/id/bundle/filter controls, readForensicParams, proof verify now queries then calls forensic.verify with returned records/start hash, export includes bundle/filter params, and static asset coverage.

Phase 28 — Merge functions for optimistic concurrency (R07)¶

• [x] T270: Conflict model (write-write, constraint, dependency) + detection hooks. Evidence: merge.apply handles expected_etag, min_causality, primary-key mismatch, and non-null constraint failures; focused tests cover conflict and non-null rejection paths.
• [x] T271: Merge function registry (Wasm) + capability model ("values-only" access). Evidence: merge.wasm.*, persisted merge_wasm_registry.json v1, validate_merge_wasm_policy, and executable values-only scalar Wasm merge modules.
• [x] T272: SkeinQL merge.register / merge.apply + SQL compat hook (If-Match). Evidence: typed SkeinQL params/results, RPC dispatch/capability advertising, ETag/min-causality merge guards, and merge_apply_wasm_policy_executes_rpc.
• [x] T273: Offline write queue format (client SDK spec) + merge result handling. Evidence: docs/OFFLINE_WRITE_QUEUE.md plus crates/skeindb-skeinql/tests/offline_queue_roundtrip.rs.
• [x] T274: Safety tests: cancellation + deterministic merges. Evidence: merge_apply_wasm_policy_cancels_non_terminating_module, merge_apply_wasm_policy_executes_values_only_module, and cluster_rpc.rs::r07_merge_conflict_resolution_deterministic.
• [x] T275: Bench: conflict rate + resolution success on example workloads. Evidence: read-only merge.evaluate returns skein.merge.evaluate.v1 with conflict/resolution rates, mean/p95 timing, and per-case results.
• [x] T276: SkeinAdmin "Merge rules" page. Evidence: Merge & CRDT panel now sends typed apply/register/simulate/evaluate/Wasm payloads and skeinadmin_merge_panel_exposes_r07_hardening_controls locks the controls.

Phase 29 — Incremental view maintenance (R08)¶

• [x] T280: view.create SkeinQL method with persisted definition (SkeinIR). Evidence: view.create, views.json format v2, restart persistence, and view_dependency_usage_persists_in_views_json.
• [x] T281: Dependency graph extension: view → base table deps at column granularity. Evidence: dependency objects include columns, projection_columns, predicate_columns, group_by_columns, and view.explain_deps traverses direct/transitive graph edges.
• [x] T282: Delta derivation for a restricted operator set (filter, project, group-by). Evidence: restricted single-table filter/project views, grouped aggregate plans for grouped columns plus COUNT/SUM/AVG/MIN/MAX, and grouped source-row persistence.
• [x] T283: Incremental refresh pipeline (apply deltas from CDC stream). Evidence: refresh_view_incremental, refresh_grouped_view_incremental, change-log driven stale marking, touched-row/touched-group recompute, and restart coverage.
• [x] T284: Cost-based switch: incremental vs full recompute. Evidence: view.refresh mode:"auto", view_refresh_stats, should_refresh_view_full, and view_grouped_auto_refresh_prefers_full_for_wide_change_sets.
• [x] T285: Correctness oracle: compare incremental vs recompute on random workloads. Evidence: view.evaluate compares cloned incremental and full refresh results and r08_view_correctness_oracle_random_workload_matches_full_recompute runs a deterministic pseudo-random workload.
• [x] T286: Bench: view maintenance overhead + query speedups. Evidence: view.evaluate returns mean incremental/full nanoseconds, speedup-vs-full, pending changes, touched primary keys, and recommended mode.
• [x] T287: SkeinAdmin “Views” page (status, refresh, explain deps). Evidence: Views panel now exposes refresh mode, evaluate iterations, view.evaluate, status, drop, dependency summaries, RPC templates, and static asset coverage.

Phase 30 — HTTP/3 / QUIC-native protocol (R09)¶

• [x] T290: Protocol sketch: SkeinQL-over-QUIC framing + stream mapping. Evidence: docs/TRANSPORT_QUIC.md defines length-prefixed JSON frames, one request/response per bidirectional stream, envelope metadata, and stream mapping.
• [x] T291: Implement server prototype with a QUIC library (feature-flag). Evidence: skeindb serve --quic --quic-cert --quic-key, Quinn integration, transport.capabilities, and quic_rpc_ping_roundtrip.
• [x] T292: Prepared query handles over QUIC streams (read-only first). Evidence: quic_prepared_query_roundtrip prepares a query, executes it on a new QUIC stream, and verifies result rows.
• [x] T293: 0-RTT safety rules (no writes in 0-RTT by default). Evidence: QUIC metadata rtt:"0rtt" is documented and quic_zero_rtt_rejects_write locks the read-only guard.
• [x] T294: Bench: p99 latency under concurrency vs HTTP/2 and MySQL/TCP. Evidence: skeindb transport-bench, transport_bench_parses_flags, latency_stats_computes_percentiles, and transport_bench_reports_http2_quic_and_mysql.
• [x] T295: Connection migration test harness (simulated IP change). Evidence: quic_connection_migration_rebind and r09_quic_concurrent_multi_stream_rpcs rebind the client UDP socket and verify continued RPC success.

Phase 31 — Vector embeddings as first-class ValueIDs (R10)¶

• [x] T300: Add ValueKind::Embedding and typed literal support in SkeinQL. Evidence: ValueKind::Embedding, Lit::Embedding { dims, v, model }, embedding_literal_roundtrip, and embedding persistence through value_store_item.
• [x] T301: LSH bucket + content hash ValueID scheme (exact + approximate id). Evidence: embedding_lsh_bucket, embedding_value_id, value_store_item using ValueKind::Embedding, and embedding_value_id_combines_lsh_bucket_and_content_hash.
• [x] T302: ANN search operator (bucket filter + distance refine) (baseline). Evidence: vector.search, HNSW HnswIndex, LSH fallback/refinement, r10_hnsw_index_basic, vector_prefilter_candidates_match_bucket, and quic_vector_search_roundtrip.
• [x] T303: Hybrid query: filter predicates + ANN order-by. Evidence: vector filters in vector.search, vector.cosine / vector.dot / vector.l2 order expressions, and query_select_orders_by_vector_similarity.
• [x] T304: Dependency tracking for embedding-derived queries (invalidate on source change). Evidence: vector.search.cache returns table-version dependencies, ETags, not_modified, and V2 causality tokens; ordinary data.* writes invalidate stale HNSW graphs; vector.insert emits prepared-query SSE invalidations; covered by vector_search_cache_metadata_tracks_source_table_changes and vector_search_cache_invalidates_after_vector_insert_rpc.
• [x] T305: Bench harness: recall/latency vs baseline index. Evidence: vector.benchmark, VectorBenchmarkResult latency percentiles, exact-vs-HNSW recall@k, vector_benchmark_reports_recall_and_latency, and quic_vector_search_roundtrip live RPC coverage.
• [x] T306: Example app: small RAG retrieval pipeline. Evidence: samples/vector_rag_pipeline.py seeds deterministic chunks, uses vector.insert / vector.search, assembles a grounded prompt, and exposes a no-server --self-test; docs/tutorials/vector-rag.md, the generated docs guide page, and sample_assets.rs / docs_site_assets.rs lock the tutorial and sample flow.
• [x] T307: SkeinAdmin “Embeddings” page (index status + query playground). Evidence: the Vector Search panel exposes DB/table/column/vector/PK/filter controls plus Search/Benchmark/Insert/Index Status actions, and skeinadmin_vectors_panel_exposes_embedding_query_playground locks the UI wiring.

Phase 32 — Natural language to SkeinQL with verification (R12)¶

• [x] T310: NL→SkeinQL prompt+schema packaging format (offline first). Evidence: AiNlPromptPackage, build_nl_prompt, ai.nl.translate, and ai_nl_translate_packages_schema / ai_nl_translate_explain_execute_roundtrip.
• [x] T311: Query explanation generator from dependency sets + planner info. Evidence: ai.nl.explain, dependencies_for_query, ai_nl_preview, and explanation fields for tables/projection/filters/order/limit/deps.
• [x] T312: Verification UI flow: explanation + sample rows + approval gate. Evidence: SkeinAdmin NL Lab query JSON / preview / approval-token controls plus ai_nl_translate_explain_execute_rpc_roundtrip.
• [x] T313: Safety policy: forbid writes unless explicit confirmation token. Evidence: the SkeinQL Query shape is read-query-only for this surface, ai.nl.explain only accepts SELECT, ai.nl.execute recomputes the approval token from query+args+deps, and tampered-query execution is rejected in ai_nl_translate_explain_execute_roundtrip.
• [x] T314: Evaluation harness: adapted text-to-SQL benchmarks (execution match). Evidence: skeindb nl-eval, NlEvalReport.execution_matches, eval_examples_exact_and_exec_match, and eval_examples_uses_rule_translation_for_execution_match.
• [x] T315: Iterative refinement protocol (user feedback loop). Evidence: prompt packages include stable fingerprint, editable generated query JSON, explicit args, preview re-explain, and reapproval before execution in SkeinAdmin.
• [x] T316: SkeinAdmin “NL Query” page (experimental). Evidence: NL Lab (R11-R12) panel wires ai.nl.translate, ai.nl.explain, and ai.nl.execute with approval-token-gated execution.

Phase 33 — Conflict-free schema evolution (R15)¶

• [x] T320: Schema version tagging in MVCC row versions. Evidence: RowEntry.schema_version / RowEntryDisk.schema_version, v3 table-row payloads with legacy v2 normalization on load, merge/apply row stamping, and schema_version_tags_row_entries_and_normalizes_legacy_rows.
• [x] T321: Concurrent schema changes protocol (add column/index) + conflict detection
• [x] T322: Query execution across schema heterogeneity (safe conversions). Evidence: schema-aware row materialization now adapts heterogeneous legacy rows via MySQL-compatible column defaults or NULL across batch/non-batch selects, keyed reads, joins, and query_select_adapts_legacy_rows_to_schema_defaults.
• [x] T323: Schema merge algorithm + roll-forward/rollback rules. Evidence: schema.apply_merge now rolls forward eligible proposals, marks deterministic losers rejected, returns rolled_back conflict details, and clears resolved losers from later schema.merge_status output in schema_evolution_merges_concurrent_column_and_index_changes and r15_schema_evolution_concurrent_column_and_index_changes.
• [x] T324: Migration assistant: show divergence + propose resolution. Evidence: schema.merge_status now returns a structured resolution plan with roll_forward, rollback, and wait actions plus caller-facing suggestions in schema_merge_status_proposes_resolution_actions and r15_schema_evolution_concurrent_column_and_index_changes.
• [x] T325: Rolling-deploy simulation harness. Evidence: schema.simulate_rollout now projects prepare/mixed/steady-state rollout stages, per-wave upgraded/legacy node counts, and legacy-row adaptation notes in schema_simulate_rollout_reports_mixed_version_waves and r15_schema_evolution_concurrent_column_and_index_changes.
• [x] T326: SkeinAdmin “Schema evolution” page. Evidence: the Schema panel now exposes typed schema.propose_change, schema.merge_status, schema.simulate_rollout, and schema.apply_merge controls plus focused asset coverage in skeinadmin_schema_panel_exposes_r15_evolution_controls.

Extensions to existing phases (inline patches)¶

The following additions extend existing phases in docs/PROJECT_BACKLOG.md.

Extend Phase 6 — Causal ETag chains (R13)¶

• [x] T064: Define causal ETag format (compressed dependencies / vector-clock hybrid). Evidence: the runtime emits vector_clock_v2 tokens, CAUSALITY_FORMAT_V2 is covered by r13_vector_clock_causality, and the wire format is documented in docs/ETAG_VALIDATORS.md / docs/SKEINQL.md.
• [x] T065: min_causality request field + response causality propagation rules. Evidence: query.select, query.execute_prepared, merge.apply, and vector.search accept min_causality and return causality; covered by query_select_min_causality_enforced and query_execute_prepared_honors_causal_cache_validators.
• [x] T066: Replication propagates causality metadata (no total order required). Evidence: replicated writes ship x-skeindb-replication-causality, replicas merge the applied watermark into cluster.status / stats.snapshot.cluster.replication, and replicated_writes_include_causality_header plus cluster_replication_ships_schema_and_rows lock the behavior.
• [x] T067: Cache interaction tests (If-None-Match with causal validators). Evidence: query_select_min_causality_enforced and query_execute_prepared_honors_causal_cache_validators verify if_none_match + min_causality on satisfied and unsatisfied dependency floors.

Extend Phase 7 — Delta topology optimization (R03)¶

• [x] T073: Implement periodic full snapshots for deltas (K-depth policy). Evidence: DeltaPolicy.snapshot_interval, ValueStore::put_with_delta, and delta_snapshot_interval_enforces_raw.
• [x] T074: Skip-pointer (skip-list) delta chain encoding. Evidence: SkipPatch, geometric build_skip_patches, materialize_with_trace, and skip_patches_reduce_steps.
• [x] T075: Compaction-time topology restructuring policy. Evidence: ValueStore::compact_deltas, DeltaCompactionReport, and delta_compaction_rewrites_deep_chains.
• [x] T076: Bench: reconstruction latency vs write amplification. Evidence: ValueStore::delta_benchmark() p50/p99/p99.9 steps, topology byte/savings metrics, and topology_analysis() depth/fanout reports.

Extend Phase 8 — Wasm query operators (R19)¶

• [x] T084: Wasm operator ABI (columnar batches) + data interchange format. Evidence: docs/WASM_OPERATORS.md, wasm.plan.compile/run, wasm_batch_v1 result format, wasm.plan.inspect, and engine::tests::wasm_plan_compile_and_run.
• [x] T085: Compile a restricted plan subset to Wasm (filter/project). Evidence: fixed-width non-null u64/bool filter/project plans now emit generated_filter_project_v1 artifacts with embedded module metadata, wasm.plan.run executes those artifacts through Wasmtime while unsupported plans fall back to host_interpreted_v1, and focused coverage lives in engine::tests::wasm_plan_compile_and_run, engine::tests::wasm_plan_compile_falls_back_for_unsupported_types, server::tests::wasm_plan_compile_run_rpc, server::tests::wasm_plan_run_batch_rpc, and tests/quic_rpc.rs::quic_wasm_plan_run_batch.
• [x] T086: Wasm SIMD exploration + perf tests. Evidence: wasm.plan.perf_report, WasmPlanPerfReportResult, latency min/p50/p95/p99/mean stats for host and generated execution, output-parity checks, explicit WasmPlanSimdExploration candidate/enabled/notes fields, and focused engine/RPC tests. This is an exploration baseline; supports_simd remains false until a future production SIMD-lowered codegen path exists.
• [x] T087: Edge runtime packaging (ship plan artifact). Evidence: wasm.plan.edge_package now emits standalone_execution metadata and a packaged JavaScript runner with runSkeinWasmPlanEdge for generated artifacts, local skein.wasm.batch.v1 encode/decode helpers, WebAssembly.instantiate execution for embedded generated_filter_project_v1 modules, and runSkeinWasmPlanHost fallback for host_interpreted_v1 artifacts.

Extend Phase 10 — Adaptive row/column execution (R02)¶

• [x] T103: Column snapshot cost model (build vs benefit). Evidence: SnapshotManager::next_plan already compares snapshot build cost against projected row-scan savings, and observe_and_plan_snapshot now feeds live table row counts so selective probes do not underestimate full snapshot build cost. Covered by engine::tests::snapshot_cost_model_uses_live_table_rows_for_selective_queries.
• [x] T104: Query pattern detector for hot projections. Evidence: QueryPatternTracker now retains a bounded per-table hot set of normalized column patterns, and only replaces colder projections when a hotter candidate arrives. Covered by engine::tests::query_pattern_tracker_retains_hot_projections and engine::tests::pattern_tracker_and_cost_model.
• [x] T105: Dependency-driven incremental refresh/invalidation for snapshots. Evidence: schema-version bumps now preserve unaffected snapshots, column renames rewrite dependent snapshot/pattern metadata, and column drops invalidate only the dependent snapshots/patterns. Covered by engine::tests::snapshot_dependencies_preserve_unrelated_snapshot_on_drop_column.
• [x] T106: Adaptive controller (online materialization decisions). Evidence: SnapshotManager::next_plan now compares candidate materializations against the best active covering snapshot instead of treating every covering snapshot as a permanent blocker, allowing narrower hot projections to replace broader ones when the online benefit stays positive. Covered by engine::tests::adaptive_snapshot_controller_replaces_broad_covering_snapshot_when_shifted.

Extend Phase 11 — Intent inference for migration (R17)¶

• [x] T114: Pattern library for common MySQL idioms (pagination, polling, soft deletes). Evidence: detect_migration_intents, detect_pagination_signal, detect_polling_signal, detect_soft_delete_signal, hierarchy/EXISTS/COALESCE detectors, and focused migration_intent_report_* tests.
• [x] T115: Sequence-level intent detection (multi-query patterns). Evidence: detect_polling_signal, increasing-value correlation in polling_values, persisted intent_history, window_ms filtering, and migration_intent_report_detects_polling_and_soft_delete.
• [x] T116: Intent → SkeinQL mapping (cursor API, CDC subscribe, etc.). Evidence: rewrite_preview_from_suggestion, rewrite_snippets_for_intent, migration.rewrite_preview, and rewrite tests for pagination, EXISTS, self-join hierarchy, and recursive CTEs.
• [x] T117: SkeinAdmin “Migration assistant” page. Evidence: the Migration (R17) panel wires migration.intent_report, migration.rewrite_preview, and migration.report_export, renders rewrite cards, and exports migration reports from web/skeinadmin/src/main.js.
• [x] T118: Offline report exporter (JSON + markdown). Evidence: migration.report_export, MigrationReportExportResult.report_json, Markdown rendering in migration_report_markdown, and migration_report_export_contains_json_and_markdown.

Extend Phase 18 — Index synthesis from dependency analysis (R16)¶

• [x] T175: Dependency capture: predicate columns + range shapes + order-by needs. Evidence: AdvisorIndexDependencyCapture is returned on advisor.index_synthesize suggestions with predicate/equality/range/order/group/join/projection columns and range_shape; covered by index_advisor_synthesizes_candidate and index_advisor_extracts_group_and_like.
• [x] T176: Candidate generator for covering + composite indexes (from deps). Evidence: advisor_candidate_columns_from_dependency builds equality/join/range/group/order composite keys, advisor_covering_include_from_dependency emits projection-derived include columns, and index_advisor_generates_composite_and_covering_from_dependencies locks the behavior.
• [x] T177: Cost/benefit model includes write overhead + compaction overhead. Evidence: AdvisorIndexCostEstimate reports read benefit, write overhead, compaction overhead, write pressure, key/include width, and net score; covered by index_advisor_cost_model_includes_write_and_compaction_overhead.
• [x] T178: Index retirement (unused) + safety rules. Evidence: advisor.retire_unused, dry-run/stale-signal safety, action retire history entries, latest-action suppression rules, and retirement tests for inactive vs recently used advisor indexes.
• [x] T179: Evaluation harness: adaptation after workload shifts. Evidence: advisor.evaluate returns skein.advisor.evaluate.v1 with phased top-suggestion convergence metrics; covered by advisor_evaluate_reports_shift_convergence, advisor_evaluate_roundtrip_reports_shift_convergence, and r16_index_advisor_evaluate_reports_shift_convergence.

Extend Phase 19 — Edge replay bundles + performance replay (R14, R18)¶

• [x] T185: Replay bundle redaction policies (privacy-safe export). Evidence: MaintenanceReplayExportParams.redaction, optional ReplayBundle.redaction, hash_pk / drop_pk primary-key redaction before checksums, SkeinAdmin/CLI controls, and replay import/run coverage for redacted bundles.
• [x] T186: Geo-distributed “bundle windows” + routing rules (bounded staleness). Evidence: edge.bundle.apply now preserves disjoint coverage windows per table while merging adjacent/overlapping ranges only, and edge.bundle.status rejects bounded-staleness routes with reason coverage_gap when contiguous coverage is incomplete; covered by edge_bundle_status_detects_coverage_gap, edge_bundle_status_reports_coverage_gap, and r14_edge_bundle_gap_blocks_bounded_staleness_route.
• [x] T187: Performance bundle extensions (LSM state, cache warm hints, timing annotations). Evidence: optional ReplayBundle.performance, ReplayBundlePerformanceProfile, storage/cache/timing sections, checksum validation, and replay_bundle_export_import_run_roundtrip.
• [x] T188: Deterministic performance replay runner + variance report. Evidence: maintenance.replay.run now rehydrates captured select/patch cache counts inside the replay workspace, computes replay-run checksum parity over reconstructable snapshot state, and keeps raw disk/WAL/cache/timing deltas in performance_report; covered by replay_bundle_run_rehydrates_cache_hints, maintenance_replay_run_rehydrates_cache_hints, and t188_replay_run_rehydrates_cache_hints.
• [x] T189: Regression CI harness: compare latency distributions across commits. Evidence: skeindb replay run --json --out, skeindb replay compare --baseline --candidate, threshold flags for p95/p99/span/storage/cache-hot-table deltas, JSON comparison reports, non-zero exit on regressions, and focused CLI tests.

Extend Phase 21 — Energy-aware compaction (R20)¶

• [x] T204: Energy model instrumentation (CPU + IO estimate; optional external signals). Evidence: CompactionEnergyConfig, CompactionEnergyRuntime, estimate_compaction_energy, and stats.snapshot.compaction.scheduler.energy.
• [x] T205: Constrained scheduler (energy minimization subject to latency/space bounds). Evidence: CompactionPolicyKind::EnergyAware, slack/constraint scoring in collect_compaction_runtime, and safe-mode override preserving hard L0 limits.
• [x] T206: External signal integration (battery/plugged, time-of-use pricing hooks). Evidence: maintenance.compaction.set_policy accepts external_signals, persists compaction.energy.*, and SkeinAdmin exposes power/price/carbon controls.
• [x] T207: Evaluation harness: energy vs p99 latency tradeoffs. Evidence: eval/compaction_scheduler_dashboard.py compares energy_aware with fixed/workload policies and emits energy score plus p99 latency summaries.

Extend Phase 22 — LLM-assisted semantic autoparameterization (R11)¶

• [x] T215: Label schema for “semantic constants” vs parameterizable literals. Evidence: ai.autoparam.label_schema exposes skein.ai.autoparam.label_schema.v1 with literal context fields, label result fields, confidence bounds, and explicit parameterize, semantic_constant, and unknown cache-key policies.
• [x] T216: Pluggable classifier interface (offline model first). Evidence: ai.autoparam.classifiers exposes the supported classifier catalog, offline_rules_v1 is selectable through classifier on classify/analyze requests, and unsupported classifier names return an error.
• [x] T217: Feedback loop: cache misses trigger reclassification. Evidence: ai.autoparam.feedback accepts cache_event: "plan_cache_miss", re-runs the selected classifier, records cached_before/reclassified, and accumulates per-fingerprint miss and reclassification counts.
• [x] T218: Metrics: plan-cache hit rate vs classifier overhead. Evidence: ai.autoparam.metrics reports plan-cache hit/miss rate together with classifier invocation/latency counters and feedback reclassification totals.

Edit this page on GitHub  ·  View source