[BUG] Python extension's venv activation is typed into Claude's prompt input instead of a shell

Refiling per the stale-close instruction on #80657 ("Please open a new issue if this is still relevant"). It is still relevant, it still reproduces on current versions, and I have a repro that does not depend on any user shell configuration — which I think is why the original report was easy to set aside.

## Summary

With `claudeCode.useTerminal: true`, the Claude Code extension launches `claude` in a VS Code integrated terminal. Independently, the Python extension (`ms-python.python`) with `python.terminal.activateEnvironment: true` — its default — calls `terminal.sendText()` on newly created terminals to activate the workspace's selected interpreter. In a workspace with a `.venv`, that text is `source /path/to/project/.venv/bin/activate`.

The two race, and Claude Code loses. By the time the Python extension sends its activation text, `claude` already owns the TTY, so the bytes are read by the TUI's stdin handler and rendered into the prompt buffer instead of being executed by a shell. The result is the worst of both outcomes:

- The venv is **never activated** — the command never reaches a shell.
- Every new Claude window opens with `source /Users///.venv/bin/activate` sitting unsent in the input box, cursor at the end, to be backspaced away before the session can be used.

Both extensions are at stock defaults. No shell rc file is involved.

## Repro

1. macOS, VS Code, a Python project containing a `.venv`.
2. Select the `.venv` interpreter for the workspace. Leave `python.terminal.activateEnvironment` at its default (`true`).
3. Set `claudeCode.useTerminal: true`.
4. Open a new Claude Code window in that workspace. The message input is pre-filled with the activation command.
5. Set `claudeCode.useTerminal: false` and reload the window. The input comes up empty.

Step 5 is the control, and it confirms the mechanism: with no terminal, the Python extension has nothing to `sendText` into, and the pollution disappears entirely.

## Expected

The input box starts empty. Anything arriving on stdin before the TUI is ready for typed input should be discarded rather than rendered into the prompt.

## Workarounds, both with a cost

`python.terminal.activateEnvironment: false` at workspace level stops it, but also disables activation in ordinary integrated terminals in that workspace, where `sendText` works correctly and is genuinely wanted.

`claudeCode.useTerminal: false` also stops it, and is the better fix in isolation, but costs the status line — context-window and cost visibility are not exposed in the native UI (#92853).

## Suggested direction

The general fix is on the Claude Code side: drain or ignore stdin received before the TUI is interactive. That covers this case, the `~/.zshenv` variant in #80657, and any other extension or shell hook that writes to the terminal at startup.

There is also a narrower interaction worth fixing regardless, which I filed separately as #92887: `claudeCode.usePythonEnvironment` (default `true`) is only honored on the SDK spawn path used by the native UI, and silently no-ops under `useTerminal: true`. I verified that by reading the extension bundle and by A/B-ing the `claude` process environment across both modes. If the terminal path contributed the environment the same way — via `EnvironmentVariableCollection`, which sets variables without writing to the terminal — the venv would actually be activated and the Python extension's `sendText` would be redundant here rather than harmful.

Taken together the two defects compound: terminal mode disables the activation mechanism that works, and exposes the one that doesn't.

## Environment

- Claude Code CLI 2.1.252, VS Code extension 2.1.263 (darwin-arm64)
- `ms-python.python` 2026.4.0
- macOS 26.6.2 (Apple silicon), zsh
- Reproduces consistently on every new Claude window in the affected workspace

## Related

- #80657 — the original report, closed as stale with an instruction to refile. Additional process-level detail is in https://github.com/anthropics/claude-code/issues/80657#issuecomment-5586877297.
- #92887 — `usePythonEnvironment` no-op under `useTerminal`, same trigger.
- microsoft/vscode-python#25744 — same root class from the extension side (activation text leaking into stdin of a process that already owns the TTY), scoped to tasks on Windows.

Reproduce with a Python workspace containing .venv, python.terminal.activateEnvironment:true, and claudeCode.useTerminal:true. Start by tracing the terminal launch path for Claude Code and where the TUI begins accepting stdin. Done means early activation text is not rendered into Claude's prompt, while the input stays empty in the reported control case.

Vom Indexierungsmodell aus dem Issue-Text verfasst.