Port bundled CPython and ordered CodeMode into Maple Agent - #897

Adds the normal permission-controlled python_code tool to Maple Agent: bundled CPython, persistent task namespaces, a continuous asyncio loop, Reset/lifecycle fencing, four retained worker slots, and up to 32 permitted Python calls executed in submission order. Other tools retain concurrency.

Rebased onto master cd4218f5. Integration retains the current maple-agent-app package / maple-agent executable, Popup menus, Active/Settled/Archived task states, external-provider admission and approval behavior, registry SDK 4.0.1, local proxy 0.4.1, and CUA 0.28.0. Settling retains Python state; archiving retires it. The component commit hook prepares the native Python fixture before running tests.

Goose remains at master's 1.53.0 package version, with the focused ordered-call patch at 19605eb50df1b27ab8f7ddd9c4e5dcbba19ec333, exactly one commit above master's previous pin. Advertised-tool/external-dispatch guards are preserved. Maple continues to select the legacy execution loop and existing live approval router. Tests explicitly select both Goose loops and use the newer approval API only in their isolated harness.

File-workflow improvements include external package installation guidance that keeps packages outside the interpreter bundle, documented Stop grace-period effects and spawned-pool requirements, bounded filesystem/syntax diagnostics, and adjacent-write coalescing that preserves stream/execution identity and output limits. The 5,000-line regression retains all 23,890 bytes.

The October 4 hosted run exposed two branch-introduced CI defects, repaired in d5eb11c5 and 4f8603d8: archive checksum/output encoding on Windows and the Intel shell's unsupported Nixpkgs/Bash startup environment. Checksums now use explicit UTF-8/LF; limited consoles receive an escaped path. Intel alone uses a separately locked supported Nixpkgs 26.05 input, retaining Python 3.13.15 and Rust 1.98.0. The Intel lane bootstraps Bash from that same lock entry so nix develop cannot fall back to incompatible system Bash. Other platform input pins and shell derivation identities are unchanged.

Validation:

• At d5eb11c5, full pinned macOS ARM64 just ci through the repository hook passed: four strict Clippy configurations, warning-denied workspace debug build, 964 workspace tests (five existing ignored), 36 headless tests, 24 worker tests plus one Windows-only skip, and seven packaging tests. Root Nix checks passed.
• The Windows encoding regression performs actual cp1252 writes/stdout on any host; both archive formats and exact UTF-8 checksum bytes pass. Native Windows worker/packaging and Linux's full feature matrix passed in run 37234348383. That run exposed Intel's additional system-Bash fallback problem, which a focused fixture reproduces and the pinned bootstrap fixes. Workflow lint, 371 repository CI tests, and root Nix checks passed for 4f8603d8.
• On final head 4f8603d8, run 37235676344 has a fully passing Intel job: dependency verification, Python worker/packaging, Rust worker tests, and offline Unicode-path relocation without Python on PATH. Windows worker/packaging passed again. ARM64 failed with the same CUA vendoring defect described below. As of 2026-10-04 21:36 UTC, the Windows/macOS ARM64 workspace jobs and Linux x64 matrix are still running; the entire hosted run is not complete or green.
• Before these CI-only fixes, Linux x64 and macOS ARM64 passed run 37186805126. Four native Python ordering cases passed under each Goose loop locally; these remain isolated scheduler tests, not authenticated Maple approval/resume proof.
• The unchanged Goose patch passed 13 focused ordering cases, 96 state-machine cases, and 21 public-loop cases. Focused Clippy used one command-line allowance for two existing result_large_err warnings in unchanged RMCP adapters; Maple's full Clippy gate used no allowances.
• Historical artifact evidence belongs to 809fe4f4: signed debug app, native worker, strict ad hoc signature verification, Unicode/spaced-path relocation, and offline external-wheel/imported-module spawn-pool checks. All 1,843 app entries stayed unchanged during dependency smoke. That app was not rebuilt for the CI-only fixes.

Remaining CI blocker: the new Linux ARM64 pure Nix lane exposes an existing CUA packaging defect. CUA core includes skill documents outside its crate directory, but Nix Cargo vendoring copies only the crate. The same CUA pin/hash, unconditional includes, and relevant package settings exist on master. The branch failed there before Python installation; a separate master pure build was not executed. Ordinary master CI does not exercise this lane. The fix is deferred rather than weakening the retained-interpreter check or broadening this feature into CUA packaging changes.

Authenticated GUI/ACP/live-model workflows, installed Linux Nix closure, browser-download Gatekeeper behavior, and release signing remain unverified. Final-value repr failure behavior and runtime-unavailable wording remain deferred. This stays Draft; original PRs #2/#69 remain untouched. No merge or release is included.