[t-kalinowski]

First evaluation currently pays for Python initialization and the default SQL connection even when neither is used. This change initializes them at their first consumer while retaining eager configured startup programs and native R startup.

r.enabled: false skips R discovery, preparation, and initialization, including on hosts with R installed. Python-only and Python/SQL sessions support managed or existing environments; an existing environment with R disabled no longer needs a retained preparation process. The documented Python/SQL configuration uses r.enabled: false, python: .venv, and languages: [python, sql].

Internally, disabled R package preparation skips an unused reticulate bootstrap probe. Managed Python reports its interpreter and prepares its optional Matplotlib cache in one child process. Matplotlib and DuckDB preparation reuse successful receipts only while the runtime, packages, preparation source, and cache identities still match. Small R preparation children omit default packages. Interpreter selection, startup-hook ordering, callback ownership, cancellation, and worker retirement have public regression coverage. Setup and restart regressions record progress only after their interpreter setup checkpoints, preserving exact output and retirement assertions. Transcript collectors also retain the final [done] marker when control-and-cell output spans replies; a gated native R startup regression and related restart continuations verify this completion contract.

On Windows, Console updates the shared C runtime’s environment view when Python initializes and when R startup hints are restored. This preserves Unicode startup-hook paths and keeps the managed environment selection visible to R after native startup. Native startup and interrupt fixtures select Python-first or R-first initialization explicitly. Portable fixtures compare interpreter-directory identity and share exact discovery transcripts for their execution mode.

On Windows, a Python startup hook that starts R cannot initialize reticulate from an R profile before Console’s adapter exists. Reticulate does not detect the running Python interpreter in this path. A native Windows acceptance case preserves the complete diagnostic; the macOS/Linux case continues to assert successful nested startup. The native and transcript acceptance cases share a dependency-light public startup fixture, so native acceptance does not import the transcript renderer or its YAML dependencies.

The benchmark report includes the harness, all timed samples, runtime versions, and binary hashes. On the measured macOS host with warm dependency caches, direct R startup with an existing Python environment and no uv fell from 1,257 to 373 ms; managed Python-only startup fell from 635 to 477 ms. Existing Python without DuckDB showed no consistent improvement. The fastest measured R configuration still takes about 1.44 times plain R, so the 10–20% overhead target remains unmet.

Local macOS validation covers all nine affected public cases (15 executions), the final shared-deadline startup case, and all 23 MCP client self-tests on identical client and collector sources. Formatting and scripts/check passed; the final fixture-only adjustments were validated by their focused public cases.

Broader local macOS coverage includes the runtime-bootstrap and progress suites (22 cases, 41 executions), 60 focused runtime cases (107 executions), all 36 website checks, and both POSIX nested-startup cases in direct and Seatbelt modes. The public nested-startup scenario also passed with Python site packages disabled. The shared-CRT synchronization function typechecks against the repository dependencies for x86_64-pc-windows-msvc; native Windows is unavailable locally.

Hosted checks for the current test-only repair are pending.