fix: compact tool results while preserving full execution evidence

[locez]

Summary

Separate tool-authored model bodies from full display/evidence artifacts so repeated model requests do not replay redundant output envelopes.

• Persist optional model text/JSON for transcript replay, session resume, and compaction, while artifact reads and checkpoint references retain full evidence.
• Return compact process status, stdout/stderr, and applicable guidance. Preserve UTF-8 text unchanged; label non-UTF-8 replacement characters explicitly and retain exact captured bytes in the full artifact instead of replaying JSON/base64 to the model.
• Compact read_text, apply_patch, and workspace failure results without dropping line ranges, truncation notices, errors, or actionable guidance.
• Unify result-content ownership and constructor validation, borrow compaction history rather than cloning full payloads, and emit recovery guidance only once per model result.
• Share process-output fields between the writer, TUI, and debug shell; preserve historical integer statuses and optional metadata, and keep malformed stream payloads inspectable as raw text.

Boundaries and compatibility

• Tools own result semantics. Generic runtime code only validates, persists, and selects the supplied bodies; provider adapters remain unchanged.
• Existing results without a separate model body keep their prior behavior. Permission review and exact evidence access continue to use full artifacts.
• Tool definitions, schemas, ordering, and stable request prefixes remain unchanged, with deterministic boundary coverage for replay and cache-prefix stability.
• Binary inspection output is not summarized by runtime. Non-UTF-8 model text is explicitly lossy; lossless captured bytes remain in artifacts.

Validation

Completed on the submitted code before creating this PR:

• git diff --check
• cargo fmt --all --check
• cargo clippy --all-targets --all-features -- -D warnings
• cargo test --all
• cargo build -p merry-cli
• Python dependency sync, Ruff, and ty
• Rebuilt the Python native extension with test-utils; 78 Python tests passed
• uv build (wheel and source distribution)

Regression coverage includes binary stdout/stderr at multiple sizes, output truncation and exceptional statuses, duplicate guidance, byte-exact artifacts, save/resume and checkpoint readback, full permission-review evidence, atomic rejection of invalid results, and TUI legacy/malformed-output handling.