fix(read-text): align read limits, guidance, and output previews

[locez]

Summary

• Make successful read_text output honor show_successful_command_output in both live sessions and transcript replay, without hiding failure diagnostics or changing runtime artifacts/model-visible results.
• Reuse one CLI preview and render-time visibility path for reads and processes. Retain at most five nonempty lines of 180 Unicode scalar values each, preserving character, line, and upstream truncation indicators.
• Raise the default/max read window from 200 to 2,000 lines while keeping the independent 1 MiB scan/return limit.
• Consolidate read scope and content-reuse strategy in the stable coding policy: focused reads when sufficient, larger or whole-file reads when required, and no unnecessary re-reading of unchanged context.
• Let unspecified Python workspace limits inherit Rust defaults instead of duplicating numeric defaults in Python.
• Consolidate read-specific TUI tests and add end-to-end Python tests for all four native limits, partial overrides, exact boundaries, invalid values, and failure without file mutation.

Boundaries and compatibility

• Preview limits are CLI-only; the 2,000-line tool window and retained artifacts are independent of presentation truncation.
• WorkspaceLimits fields now remain None when inherited. Explicit positive values still override individual Rust defaults; Python configuration is not a resolved-default snapshot. SDK documentation and native type stubs are updated.
• The coding prompt, read tool description, and default schema change the request prefix across versions. The new prefix and tool order remain stable within a session; this does not introduce per-turn prompt/tool churn.
• No new crate, runtime state owner, or provider protocol is introduced. The read-result DTO remains unchanged.

Verification

• git diff --check
• cargo fmt --all --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test --all — 2,219 passed, 9 ignored
• Python uv sync and maturin develop --uv --features test-utils
• Ruff format/check and ty check — zero diagnostics
• Python pytest tests -q — 78 passed
• uv build — source distribution and wheel built successfully

Long-line preview regression tests fail before the preview bound is restored and pass afterward. Python limit tests use a local deterministic model fixture and execute the actual Rust workspace tools. No live-model comparison is claimed for the prompt's behavioral effects. Packaging still reports existing Cargo manifest metadata warnings.