feat(cli): add --overwrite/--no-overwrite output safety to all output commands

[xieofxie]

Summary

Output-producing commands silently overwrote an existing file/dir on a re-run — no warning, no prompt, no opt-in (#561). This adds a shared safe-overwrite control so a re-run can't destroy a previous result.

Flag name

Chose --overwrite/--no-overwrite (default --no-overwrite) over the issue body's --force/-f because it matches the dominant --x/--no-x toggle house style, is precise to the file-overwrite concern, avoids the -f collision (run --file), and won't be confused with the existing --rebuild/cache controls.

Shared helpers (utils/cli.py)

• overwrite_option() — --overwrite/--no-overwrite, default no-overwrite.
• guard_output(path, overwrite, *, label) — fail-fast existence check. Files block when they exist; directories block only when non-empty (a fresh/empty output dir doesn't false-trigger); a None path (stdout) is a no-op.

Wired into 10 commands

export, optimize, quantize, compile, config, eval, analyze, perf, catalog, run.

• Guard runs before any mkdir/cleanup/work — notably before the quantizer's destructive stale-sidecar unlink().
• Covers default-derived output paths (e.g. {stem}_qdq.onnx), not just explicit -o.
• analyze guards both --output and --optim-config.
• compile guards the resolved -o file or --output-dir (non-empty).

build intentionally excluded

build already has --rebuild plus incremental resume — without --rebuild, existing stage artifacts are reused, not clobbered (e.g. quantize stage skips when QDQ nodes are present). A non-empty-output-dir guard there would break legitimate incremental rebuilds and overlap confusingly with --rebuild.

Tests

• guard_output / overwrite_option unit tests (file/dir/empty-dir/overwrite-bypass/None/custom-label).
• Real-command wiring: catalog (file), quantize (file + default-derived path + verifies quantize_onnx is not called when blocked), compile (empty dir OK, non-empty blocked, --overwrite proceeds).
• Full tests/unit/commands/ + tests/unit/utils/test_cli.py suite green (916 + new tests).

Fixes #561

[DingmaomaoBJTU]

Overall the design is clean and the coverage is thorough — approve with two minor notes left as inline comments.

• guard_output() utility is well-scoped (non-empty dir semantics, None is a no-op, str coercion).
• Fail-fast placement is correct in the majority of commands (before model load / eval / compile).
• Test suite covers the happy path, both flag polarities, default derivation, str vs Path, custom label, and directory cases.