v0.0.3

Added

• Added local OpenAI-compatible embedding endpoint support via evo.embedding_model=local/<model>@http(s)://host[:port]/v1.
• Added CONTRIBUTING.md plus GitHub issue and pull-request templates to document the contribution flow.
• Added Python throughput plotting utilities in shinka.plots for generation runtime timelines and normalized occupancy-over-time views.
• Added a durable SQLite generation_event_log journal for async generation lifecycle debugging, including stop and persistence-failure events.
• Added regression coverage for the new Python throughput plotting helpers, including pool-slot prep, occupancy math, and legend/layout behavior.
• Added regression coverage for concurrent async completed-job persistence so multi-worker postprocessing throughput stays exercised.
• Added regression coverage for lightweight program summaries and WebUI embed-tab hydration so similarity matrices only render from fully loaded embedding data.

Changed

• Reworked async completion handling so completed scheduler jobs are detected immediately, evaluation slots are released before persistence finishes, and shutdown now waits for queued completed-job batches plus post-persistence side effects to drain.
• Moved database archive / best-program / island maintenance off the insert hot path via deferred replay hooks, while letting async writers use fresh worker-local connections and merge runtime metadata back into the shared DB state.
• Expanded pipeline timing metadata with post-evaluation queue wait, post-persistence side-effect timing, and summary statistics for end-to-end async throughput analysis.
• Tuned examples/circle_packing/shinka_long.yaml for a smaller long-run preset and ignored generated results* / shinka_scale* artifacts in the repo root.
• Renamed the local backend guide from docs/support_local_llm.md to docs/support_local_models.md and expanded it to cover local embedding backends alongside local LLMs.
• Refactored async code validation to use a shared subprocess helper across Python, Rust, Swift, JSON, and C++ validators without changing validation behavior.
• Updated examples/circle_packing/load_results.ipynb to include the new throughput plots at the bottom of the notebook.
• Updated examples/circle_packing/load_results.ipynb and examples/circle_packing/shinka_long.yaml for the latest large async circle-packing run analysis setup.
• Refined Python throughput plot legends to use compact centered panels below each subplot for cleaner notebook rendering.
• Reduced the async generation journal to high-signal failure/stop events only so persistence debugging does not add heavy hot-path overhead.

Fixed

• Fixed completion-time accounting so retried or duplicate-persisted jobs keep the original scheduler completion timestamp instead of inflating evaluation duration.
• Fixed the async job monitor to finalize cleanly once the generation target is reached, even when no jobs remain active at the polling boundary.
• Fixed high-concurrency SQLite persistence regressions by covering deferred maintenance replay, multi-writer overlap, and shutdown drain behavior with new recovery and database tests.
• Fixed async proposal scheduling so num_generations is now a hard cap on assigned proposal generations instead of launching extra gen_* attempts to compensate for failed or discarded work.
• Fixed async evaluation slot lifecycle bugs so local evaluation concurrency no longer exceeds max_evaluation_jobs through stale double-release of reassigned worker slots.
• Fixed retry-time completion accounting so successful DB retries refresh async progress before the hard generation-budget stop condition is evaluated.
• Fixed async database retry races by treating in-flight source_job_id inserts as already claimed, preventing duplicate persisted programs while timed-out writes are still finishing in worker threads.
• Fixed async resume/recovery bookkeeping so restarted runs continue from the number of persisted completed programs instead of stopping early when failed proposals or hung local evals left gaps in generation IDs.
• Fixed WebUI meta-analysis labeling so meta_*.txt snapshots are presented as meta updates / processed-count checkpoints instead of misleading generation numbers.
• Fixed duplicate-retry recovery so already-persisted jobs replay post-persistence side effects exactly once, restoring missing meta-memory updates after DB timeouts.
• Fixed SQLite persistence stability by increasing busy timeouts and the outer async DB-add timeout for long high-concurrency runs.
• Fixed Python throughput plot preparation so frames without optional metadata columns like is_island_copy, patch_name, or model_name still render correctly.
• Fixed legacy throughput accounting in both the Python plotter and WebUI Throughput tab so reused worker lanes no longer show impossible peaks like 31/20.
• Fixed the WebUI embed tab so summary-only loads or single lazily hydrated programs no longer produce a misleading 1x1 similarity matrix instead of the full run.