Add repeatable prompt and diagnostic benchmarks

[jiminu]

The existing benchmark measured shell initialization but stopped before a prompt was displayed, and did not cover status or doctor. Add optional --prompt and --diagnostics measurements to the existing base/full modes and TSV output.

• Measure first-prompt and command-to-prompt latency in an empty directory and the source repository using Python's standard-library PTY support. Numbered markers reject editing redraws; graceful shutdown releases history locks before the next sample.
• Install configuration in a temporary HOME before timing diagnostics, report the observed installation and exit codes, and fail if a sampled exit code changes. Preserve WSL platform context and bound mise configuration discovery to avoid ancestor settings affecting results.
• Include the missing history module in the benchmark fixture and document measurement scope, dependencies, and limits.

Validation:

• bash scripts/check.sh passed locally on macOS: all 21 suites, Bash/Zsh syntax, ShellCheck, and shell formatting.
• Regression checks cover numeric metric output, prompt sequencing, timeout/closed-shell behavior, isolated environment settings, and normal exit hooks. Code review completed.
• Full-mode local run: 30 first-prompt samples per scenario, 90 command samples per scenario, and 30 samples per diagnostic command. p50: first prompt 90–109 ms, command-to-prompt 31–49 ms, status 625 ms, doctor 640 ms. These are machine-specific observations, not performance budgets.
• The diagnostic fixture uses temporary pinned shell tools plus host package/PATH inventories. Its cask inventory encountered missing tap trust metadata in the isolated HOME; timings include that error path and do not represent a freshly provisioned complete environment.