0.5.3rc4 scope and validation

Rc4 integrates the independent rc3 audit and retains the published scientific coding-agent leaderboard and local/SSH workers. Minimal remains the default.

Execution and evidence repairs

  • Serialize SQLite appends before reading the chain head, compare canonical JSON for replay identity, and index campaign heads and incremental replay.

  • Account for explicitly reserved CPUs in the process CPU-time limit, preserving the existing wall-time allowance for threaded experiments. This is not CPU affinity or an aggregate MPI process-tree quota.

  • Reduce targeted worker-poll overhead while retaining complete admission and reservation reconciliation. Stop replayed uploads only after completed input checksums are verified. Preserve uncertainty after malformed transport replies.

  • Stream request usage fields from provider logs, clarify research-client errors, validate finite durations and nested required inputs before persistence, and handle extreme finite parameter ranges without intermediate overflow.

Interface and documentation

Conversation submissions capture their destination; stale navigation responses cannot replace a newer selection. Preserve newly typed drafts, guard duplicate benchmark submissions, retain appropriate disabled buttons, clear closed-dialog SSH passwords, and explain connection-save failures inline.

Visible polish adds keyboard skip links, accessible names, shared monitor/workspace colors, darker secondary text in light mode, composer help, reduced-motion behavior and bounded dialog scrolling. Research tools, Machines and Benchmarks retain their existing page layouts.

Documentation distinguishes stable and preview capabilities and records actual execution boundaries. The contribution guide’s authorship text remains explicitly a discussion draft; this release adopts no publication-authorship entitlement.

Direct reproductions

On the exact released rc3 base, eight concurrent SQLite connections retained all 288 appended events but produced an invalid hash chain. The audited implementation retained the same number of events and a valid chain. It also rejected the distinct JSON payloads true and 1 where rc3 accepted an idempotent replay.

An eight-thread computation with eight reserved CPUs and a three-second wall allowance exited 137 after 0.556 seconds under rc3. The audited launcher completed it successfully in 1.361 seconds, with 10.3284 aggregate CPU seconds. This is a bounded operating-system resource check, not a solver-throughput measurement.

Validation and limits

Independent review of the audited base passed 974 Python tests with seven optional runtime skips, Ruff and strict Sphinx. Actual Chromium checks covered five pages at 1440/390px in light/dark themes, without JavaScript errors or horizontal overflow. The workspace bundle installed in a fresh HOME without an agent/API key and without system package changes; HDF5 reading and research skill resources were available.

Release preparation also caught installed Node dependencies entering a source distribution built from a Git worktree. Explicit source-archive exclusions now remove dependencies, environments and local state; the rebuilt archive is checked independently of Git-ignore detection.

Rc4 release checks cover its own complete Python suite, DSH/MCP adapter, schema consistency, strict docs, distributions and a fresh-home installer. CI additionally qualifies the bounded WarpX CPU instrument and recorded deterministic campaigns. The GitHub release links the exact commit and workflow results.

Synthetic query, polling and log-reader measurements support narrower orchestration and memory improvements. They do not measure model token savings, solver speedups or end-to-end research success. The audit’s restricted-host failures did not reproduce on this review machine; that does not certify every restricted host.

No new long GPU/FLASH/WarpX, authenticated remote-worker or live paid-model research campaign was performed for this release. The scientific search strategy, prompts, completion rules, task pack 0.3.0 and published model measurements are unchanged. Historical reconnection wording now distinguishes excluding the exact exponent from falsifying the wider registered tolerance band; original records are retained.