fix: use with_global_system() to prevent FD leak in Jetson, Tenstorrent, and NVIDIA readers (#120)

[inureyes]

Summary

Fixes #120 - Replace per-call System::new() instances with with_global_system() to prevent file descriptor leaks in API mode (long-running metrics loop).

Problem

The NVIDIA Jetson, Tenstorrent, and NVIDIA readers were creating new System::new() instances on every get_process_info() call. In API mode, this runs every interval indefinitely, leaking /proc file descriptors each cycle.

Changes

NVIDIA Jetson (src/device/readers/nvidia_jetson.rs)

• Line 176: get_process_info() now uses with_global_system()
• Line 261: get_gpu_processes() helper now uses with_global_system()
• Removed unused use sysinfo::System; import

Tenstorrent (src/device/readers/tenstorrent.rs)

• Line 201: get_process_info() now uses with_global_system()
• Removed unused use sysinfo::System; import

NVIDIA (src/device/readers/nvidia.rs)

• Removed system: Mutex<System> field from NvidiaGpuReader struct
• Migrated get_process_info() to use with_global_system() for consistency with the rest of the codebase

Approach

Uses the with_global_system() function from src/utils/system.rs which provides a global Mutex<System> instance (GLOBAL_SYSTEM). This is the standard pattern already used by:

• AMD reader (src/device/readers/amd.rs:583-585)
• Apple Silicon reader (src/device/readers/apple_silicon_native.rs:388)
• Local collector (src/view/data_collection/local_collector.rs:304, 433)

Testing

• ✅ Code compiles successfully with cargo build
• ✅ All three readers now use the same consistent pattern as other readers in the codebase
• ✅ No functional changes to the API, only internal optimization

Closes #120

[inureyes]

Security & Performance Review

Analysis Summary

• PR Branch: fix/120-fd-leak-jetson-tenstorrent (all analysis performed here)
• Scope: changed-files with full-context cross-reference
• Languages: Rust
• Total issues: 0 actionable (1 pre-existing observation noted below)
• Critical: 0 | High: 0 | Medium: 0 | Low: 0

Review Checklist

Security

• No security vulnerabilities introduced
• No new unsafe blocks
• No new external input handling
• No hardcoded credentials or secrets
• No command injection vectors

Performance - Global System Pattern

• with_global_system() correctly applied in all three readers (nvidia.rs, nvidia_jetson.rs, tenstorrent.rs)
• No nested with_global_system() calls that could deadlock -- all uses are sequential within each call chain
• Mutex contention analysis: In local_collector.rs, GPU reader get_process_info() (which now uses with_global_system) runs concurrently with the "full process collection" task (which also uses with_global_system). This introduces contention on GLOBAL_SYSTEM, but NOT deadlock -- one task will simply wait for the other. This is acceptable and consistent with the existing AMD reader behavior.
• Lock hold duration is appropriate -- only held during process refresh and data extraction, then released
• No System::new() calls remain in any of the three changed files

Correctness

• All System::new() calls properly replaced with with_global_system()
• NVIDIA reader: system: Mutex<System> field correctly removed from struct, constructor, and usage
• Jetson reader: Both get_process_info() (line 179) and get_gpu_processes() fallback path (line 261) correctly use with_global_system()
• Tenstorrent reader: get_process_info() correctly uses with_global_system()
• use sysinfo::System; import correctly removed from all three files
• with_global_system import correctly added to all three files
• Code compiles cleanly (cargo check passes)
• No clippy warnings (cargo clippy -- -D warnings passes)
• All 19 tests pass

Code Quality

• Consistent pattern with existing AMD and Apple Silicon readers
• No unused imports
• Comments properly updated to reflect the new pattern
• Proper scoping of use sysinfo::{...} imports within functions

Pre-existing Observation (Not Introduced by This PR)

In nvidia_jetson.rs get_gpu_processes() fallback path (line 261-296): When with_global_system is used, only system.refresh_memory() is called before iterating system.processes(). Process data is NOT refreshed, so it reads whatever process state was last populated by another caller. This was the same in the original code (System::new() + refresh_memory() would have had an empty process list), so this PR does not regress behavior -- it actually slightly improves it since the global system may contain previously-refreshed process data.

Verdict

This PR is clean and ready to merge. The changes are mechanical, consistent, and correctly follow the established with_global_system() pattern. No issues found that require fixing.

[inureyes]

PR Finalization Report

Project Structure

• Project Type: Rust (Cargo)
• Test Framework: cargo test (166 lib + 169 bin + 32 integration + 19 doc tests = 386 total)
• Documentation System: Plain markdown in docs/
• Lint Tools: cargo fmt, cargo clippy (with .clippy.toml config)

Checklist

Tests

• Analyzed existing test structure (unit tests in src/, integration tests in tests/)
• Identified test coverage needs: No new tests required -- this is a pure internal refactor replacing System::new() with with_global_system() across three readers. No public API changes.
• All 386 tests passing (0 failures)

Documentation

• Reviewed documentation needs: No updates required
  • No public API changes (internal refactor only)
  • No new features or configuration options
  • No changes to CLI behavior or output

Code Quality

• cargo fmt --check: Passed (no formatting issues)
• cargo clippy -- -D warnings: Passed (no warnings)
• All unused imports (sysinfo::System) properly removed
• Consistent pattern applied across all three readers (nvidia, nvidia_jetson, tenstorrent)

Changes Verified

• src/device/readers/nvidia.rs: Removed system: Mutex<System> field, migrated get_process_info() to with_global_system()
• src/device/readers/nvidia_jetson.rs: Migrated both get_process_info() and get_gpu_processes() to with_global_system()
• src/device/readers/tenstorrent.rs: Migrated get_process_info() to with_global_system()

All three readers now match the pattern already used by AMD, Apple Silicon, and local collector readers.

Status

All checks passing. Ready for merge.