Launch managed SDK servers through the Rust runtime wrapper

[roji]

Managed out-of-process SDK connections currently launch the root Copilot SEA directly. The runtime now provides a dedicated copilot-runtime executable that loads the adjacent runtime.node, so SDK-managed subprocess connections should use that coherent wrapper/addon pair while preserving explicit external server and in-process modes.

What changed

• Update Node.js, Python, Go, .NET, Rust, and Java managed launch paths to resolve and spawn copilot-runtime (copilot-runtime.exe on Windows).
• Materialize and validate the wrapper beside runtime.node, retaining the root copilot executable only for residual and in-process compatibility.
• Preserve auxiliary runtime assets such as ripgrep, tgrep, built-in definitions, skills, and future package additions. Each SDK keeps unknown npm package entries by default while excluding known CLI/UI-only content.
• Preserve explicit CLI path, process, and server URL overrides without silently falling back to the SEA when managed wrapper artifacts are incomplete.
• Extend language-specific packaging, extraction, permissions, cache, and test coverage for the coherent runtime distribution.
• Fix a Python subprocess stderr-reader race and Java MCP stdio cwd serialization issue discovered during end-to-end validation.

Validation

Validated all six SDKs against the immutable same-tip runtime release candidate adce886fd1 across supported root SEA, in-process, and hostless wrapper transports. The final focused cache/resume gates and supported full-suite cells passed with exact artifact hash checks and no leaked runtime processes.

The auxiliary-asset follow-up adds focused staging, cache-migration, filtering, permission, and path-safety tests in all six SDKs. A real Node hostless-wrapper integration test also confirms built-in grep completes successfully through the staged ripgrep binary.

Known non-runtime exclusions remain documented from validation: hostless Node extension/factory lifecycle behavior (including the corresponding .NET extension tests), the Go invalid-model fixture, the Java literal-YAML fixture and order-dependent socket-helper contamination, and Rust forced-environment unit-test contamination.

Review notes

This is a draft while the runtime platform packages and Java classifier publication complete. The production platform package is treated as the runtime distribution source: the root copilot[.exe], selected prebuilds/<platform>/runtime.node and copilot-runtime[.exe], and retained package-relative runtime assets are staged as one coherent version. Mixed or incomplete artifact sets are rejected.

[SteveSandersonMS]

The hostless wrapper direction looks right, but this is not complete yet because runtime.node is not self-contained.

I manually tested all 48 combinations: 6 SDKs × stdio/TCP/in-process/existing URL × source/published consumption, using @github/copilot@1.0.81-9. Startup succeeded without launching the SEA in every case. However, built-in grep failed with ENOENT in every stdio/TCP case and in 10 of 12 in-process cases. Existing-URL cases passed because that runtime used the intact npm package layout.

The missing piece is preserving the hostless runtime assets relative to runtime.node. A reasonable minimal payload is:

prebuilds/<platform>/runtime.node
prebuilds/<platform>/copilot-runtime[.exe]
ripgrep/
tgrep/
definitions/
builtin/
builtin-skills/

Node-only and delivery-only content can be excluded: copilot, app.js, index.js, npm-loader.js, sea-loader.js, copilot-sdk/, sdk/, cli-native.node, preloads/, foundry-local-sdk/, pvrecorder/, voice-*.js, webview/, tree-sitter*.wasm, queries/, UI assets, metadata, and the duplicate copilot-runtime-bin.

I validated this approach on .NET in both development and after dotnet publish. Preserving ripgrep/bin/<platform>/rg fixed built-in grep for stdio, TCP, and in-process. Adding tgrep/bin/<platform>/tgrep also worked with USE_TGREP=true. I repeated the runs after deleting the SEA; all passed. In-process worked using adjacent runtime.node as the entrypoint/root hint.

Before merging, I think this PR needs to:

1. Materialize the complete hostless asset set in all six SDKs while preserving npm-relative paths.
2. Remove the SEA from published outputs and embedded bundles.
3. Make in-process resolution use the hostless runtime layout rather than retaining the SEA.
4. Add E2E coverage that asserts successful tool.execution_complete events for built-in grep/tgrep in source and published consumers. Checking startup or final assistant text is insufficient because the model can silently fall back to another tool.
5. Future enhancement: Ideally define this asset set once upstream as a manifest/archive so six SDK implementations do not drift.