The exit-marker contract reaches the model only through the pwsh tool description; the bash one is silent #2045

The two shell dialects are peers that must differ only in the shell (design §14.5 item 1, and the principle behind #2039). Their renderers now agree. Their tool descriptions — the only channel through which the model is told what a finished run looks like — do not.

Measured

pwsh_tool_v2.py:707 states the contract:

"Non-zero exits are reported as `[exit code: N]`."

bash_tool_v2.py, read in full, states it nowhere. Its description is 359 chars and mentions no marker:

"Execute a shell command and return stdout and stderr. Use for running tests, git commands,
listing files, installing packages, and other shell operations. The command runs under
`bash -c`, and the kernel confines it to the session's working directory: writes outside it
(and outside the OS temp area) are refused whatever language or subprocess attempts them."

A repo-wide search for the token confirms this is the only channel, and emrg/server/prompts/system.j2 does not mention it either:

$ grep -rn "exit code: N" emrg/ --include=*.py --include=*.j2
emrg/tools/pwsh_tool_v2.py:707:   "`[exit code: N]`. The kernel confines the run to the session's "

So a model driving the bash tool is never told that a non-zero exit arrives as [exit code: N], while the same model on Windows is. The rendering is one policy; the description of it is not.

Two sentences in the tree already claim the opposite

On the head of #2042, two places assert that bash's description carries it, and neither is true:

• emrg/tools/bash_tool_v2.py (render_result docstring): "the tool's own description promises the line only for non-zero exits";
• tests/test_pwsh_tool_v2.py (test_the_rendered_contract_is_the_one_the_description_states): "(both tool descriptions say so)".

master carries no such claim, so this arrives with #2042 — whose own subject is that the two twins must not disagree. This is the class this project already names elsewhere (tests/test_glob_tool.py: "a promise in a tool description that no test holds is how the universal claim above survived review").

Expected

1. Both descriptions state the same exit-marker contract, so the dialect's shell is the only difference the model is told about; and
2. the promise is held by a test on each dialect — the marker the description names is the marker that dialect's renderer emits for a non-zero exit, and none for a success, so the sentence cannot drift from the behaviour it promises.