0.5.3 stable scope and validation

This release brings the 0.5.3 candidate series into one stable version. Minimal remains the default; structured and frontier modes are still selectable. The installer preserves research data, provider settings, installed instruments and older program folders.

Included behavior

  • Local/SSH numerical workers with recorded job identity, CPU/RAM/GPU admission, asynchronous execution, verified artifact retrieval and interruption recovery. The Machines page supports basic SSH setup, advanced settings, agent preparation and independent 30-second availability checks. Research agents and scientific review remain on the coordinator.

  • The measured coding-agent leaderboard, task-specific time/cost ranks, visible unfinished costs and Pareto tradeoffs. Custom trials and contributed grades are separate from the maintained official edition. These tasks analyze recorded diagnostics; they do not launch new FLASH or WarpX simulations.

  • Campaign recovery, request-level usage records and native DeepSeek tool-message continuity, with bounded evidence retrieval by independent reviewers.

  • The independently audited ledger, reserved-CPU accounting, worker transport, input validation and conversation repairs described in the rc4 acceptance record.

  • Shared Conversation → Study → Evidence & review navigation. A selected study survives reloads, mode changes and monitor Back/Forward navigation. Continuation returns to the owning conversation when there is one; standalone studies remain inspectable without creating a conversation during read-only access.

  • Protection against stale study replies and duplicate continuation submission. Leaving a pending continuation preview frees the action immediately; its late response cannot reopen or replace a newer dialog, even after returning to the original route.

  • Shared logo artwork across the workspace and monitor, theme-aware SVG lockups and SVG/ICO favicons. Artwork and font attribution ship in both Python distributions and the one-command workspace bundle.

  • Workspace-first onboarding, a provider-neutral headless recipe, a shorter contribution guide and an archive for dated reports. No publication-authorship policy is adopted.

Release checks

The integrated stable source passed 1,093 Python tests with seven optional-runtime skips, and 14 DSH/MCP tests. Ruff, committed schemas, JavaScript/shell syntax, strict Sphinx and 5,728 local rendered-link/asset checks passed.

The full suite includes eight real Chromium study-navigation cases and 16 new logo cases. These cover late continuation replies, exact conversation ownership, reload/Back/Forward state, light/dark themes, desktop/tablet/mobile widths, keyboard home links and saved narrow sidebars. An additional 12 combined study, evidence and continuation views at 1440/390 px had no JavaScript errors or horizontal overflow.

The source distribution, wheel and one-command workspace bundle were built and inspected for local environments, dependencies, credential files and database state. All seven brand files/notices match the source byte-for-byte; shared navigation JavaScript is present in every distribution. The checksummed workspace bundle installed under a fresh HOME with no API key or native agent and no system-package changes. Its GUI, shared assets, HDF5 reader and packaged research skills were available. The installed package also exported the official leaderboard without making a model request.

The deterministic kinetic/PIC benchmarks, planted/PIC discovery packages, blinded search and budgeted multi-action fixtures passed. Recorded Gray–Scott, GEM and agent-role verification retained their immutable artifacts and conclusions.

The seven local skips require non-root PRoot or separately installed WarpX, atoMEC, Singularity-EOS, FLASH and WSL CUDA/openPMD runtimes. They are not passes. GitHub CI additionally exercises the bounded WarpX CPU qualification and durable analytic-to-WarpX campaign. Its exact-commit results are linked from the release PR; these checks remain separate from local verification and historical candidate runs.

Limits

This release retains the research strategy, independent scientific review gates, completion contracts and benchmark task pack 0.3.0. It introduces no new model benchmark measurements or claims of improved scientific delivery. The release checks do not substitute for a new long GPU/FLASH/WarpX research campaign.

Cooperative PRoot is not an operating-system security boundary; numerical sandbox settings do not enclose native agent host tools. Optional runtimes must be commissioned on the actual worker and execution path. Passing an execution or interface check does not establish numerical convergence or physical validity.

Standalone continuation submission still has no durable retry receipt for an ambiguous lost HTTP response. Repeatedly submitting after such a loss can create another conversation. The new dialog guards do not claim to solve that case.