MENS Subsystem LoC Audit (2026-09)

Why this exists

A deep-research mandate covering MENS (in-house Qwen fine-tuning) training infrastructure — memory sizing, quantization, Ollama support, CUDA/Metal SSOT, cloud training — explicitly asked for a “LoC audit against the codebase.” None of the three execution plans that followed the mandate produced one. This is that audit, scoped the way the original ask implies: not a raw wc -l dump, but an accounting a MENS maintainer would actually use to plan further work.

Tooling note: this repo has no tokei/cloc/scc installed and no existing LoC-counting subcommand in vox-code-audit or vox-arch-check — the closest existing tooling is the arch/god_object detector (file-size thresholds only, not a corpus-wide counter; see below). Counts here were produced with wc -l / grep -c over the exact directories this document lists, run from a clean worktree at commit 3ebc4d42c. where-things-live.md confirms [guards] loc_budget = "off" workspace-wide (see Structural Limits) — there is no per-crate LoC gate to compare against, only the file-level God Object thresholds used below.

Workspace scale, for context

Language	Scope	Lines
Rust	crates/**/*.rs	~890,600
TypeScript	crates/vox-gui/ui/src/**/*.{ts,tsx}	~59,500
Vox	examples/**/*.vox, scripts/**/*.vox	~19,000

The MENS surface

Directories counted (per the mandate’s scope):

• crates/vox-populi/src/mens/
• crates/vox-plugin-mens-candle-cuda/
• crates/vox-plugin-mens-candle-metal/
• crates/vox-quantize/
• crates/vox-ml-cli/src/commands/mens/
• crates/vox-ml-cli/src/commands/schola/ (native LoRA training worker for vox mens train; entirely MENS-related, no partial-inclusion judgment call needed)
• crates/vox-gui/ui/src/components/surfaces/Models/

Crate / path	Rust files	Lines
vox-populi/src/mens/	71	17,708
vox-plugin-mens-candle-cuda/	46	9,492
vox-plugin-mens-candle-metal/	46	8,852
vox-quantize/	10	2,530
vox-ml-cli/src/commands/mens/	37	10,872
vox-ml-cli/src/commands/schola/	7	2,327
Rust subtotal	217	51,781
vox-gui/ui/.../surfaces/Models/ (TypeScript)	—	464
MENS surface total		~52,245

That’s roughly 5.8% of workspace Rust concentrated in six crates/dirs — a meaningful, but not runaway, share for a subsystem this actively worked on this quarter.

Test-to-production ratio

The repo’s Test-First Policy (AGENTS.md) requires a test alongside every new pub fn. Measuring this by line count (not test count, which the policy already gates at commit time) answers a different question: is coverage thin even where a test technically exists?

Method: for each of the 217 Rust files, everything from the first #[cfg(test)] or mod tests marker to end-of-file is counted as test code; everything before it is production code. This undercounts files that interleave #[test] functions outside a trailing mod tests block, so treat it as a lower bound on test share.

	Lines	Share
Production (approx.)	38,325	74.0%
Test (approx., lower bound)	13,456	26.0%

A 26% test-line share across 217 files is a reasonable floor, not a red flag — but it’s an aggregate. Two files that are dedicated test suites (eval_gate/tests.rs at 921 non-blank lines, hardware/tests.rs) do the heavy lifting; a maintainer adding new MENS pub fns should not assume the aggregate ratio describes their file specifically.

Largest files vs. the God Object thresholds

The arch/god_object detector (crates/vox-code-audit/src/detectors/god_object.rs) is this repo’s actual, enforced size gate — no custom threshold invented for this audit. Its tiers, on non-blank lines, per file:

• Info > 300 lines
• Warning > 400 lines
• Error > 500 lines (Severity::Error, blocking under the default legacy CI mode)

Running that measure over the 217 MENS files:

Tier	Files
Error (>500)	26
Warning (400–500)	6
Info (300–400)	21

23 of those 26 Error-tier files are not in contracts/toestub/suppressions.v1.json — only vox-ml-cli/src/commands/mens/populi/dispatch.rs, vox-plugin-mens-candle-metal/src/candle_qlora_train/training_loop/mod.rs, and vox-plugin-mens-candle-cuda/src/candle_qlora_train/training_loop/mod.rs carry a suppression entry. The other 23 are live, unsuppressed arch/god_object Error findings today. This is worth a follow-up: either the god_object detector isn’t wired into the gate MENS code actually runs through, or these findings are landing as Warning-surfaced-but-not-blocking noise nobody has triaged, or CI’s default legacy mode genuinely fails on them and it’s being tolerated/missed. Whichever it is, it’s inconsistent with the “Error blocks” contract stated in AGENTS.md’s Local CI Gate Tiers section.

Ten largest MENS files by non-blank line count:

File	Non-blank lines
vox-populi/src/mens/tensor/preset_schema.rs	1,368
vox-plugin-mens-candle-metal/src/candle_qlora_train/mod.rs	1,195
vox-plugin-mens-candle-cuda/src/candle_qlora_train/mod.rs	1,159
vox-plugin-mens-candle-metal/src/model.rs	1,096
vox-plugin-mens-candle-cuda/src/model.rs	1,092
vox-populi/src/mens/tensor/memory_model.rs	1,046
vox-ml-cli/src/commands/schola/train/gpu.rs	1,045
vox-populi/src/mens/cloud/pipeline_dispatch.rs	1,042
vox-populi/src/mens/tensor/domain_router.rs	925
vox-ml-cli/src/commands/mens/eval_gate/tests.rs	921 (test file)

preset_schema.rs, memory_model.rs, domain_router.rs, pipeline_dispatch.rs, and gpu.rs are all production code well past the 500-line Error line and are plausible single-responsibility-violation candidates: memory_model.rs and domain_router.rs in particular sit in tensor/, a directory already split into 30+ narrow files — their size looks like a natural place decomposition stopped rather than an intentional design choice.

candle-cuda / candle-metal duplication

The defactor policy (AGENTS.md, Dependency Discipline §3) allows duplicating a helper under ~50 lines with a // vox:defactored-from <crate> <date> comment instead of taking a crate edge; anything larger is supposed to be split into a shared -types/-core crate.

Comparing the 44 filenames that exist in both vox-plugin-mens-candle-cuda/src/ and vox-plugin-mens-candle-metal/src/:

• 24 of the 44 are byte-for-byte identical between the two crates, totaling 3,191 lines (one side; ~6,380 lines counting both copies).
• Only 1 of those 24 (candle_qlora_train/oom.rs, 149 lines) carries a vox:defactored-from marker.
• The other 23 unmarked identical files range from 3 lines (hf_layout.rs) up to 720 lines (qlora_preflight.rs), with several more in the hundreds: hf_keymap.rs (399), checkpoint_state.rs (236), manifest.rs (367, 13-line diff — near-identical), merge.rs (335).

None of the large ones — qlora_preflight.rs, hf_keymap.rs, checkpoint_state.rs, merge.rs — fit the policy’s “~50 lines” defactor carve-out; at that size the policy calls for a shared -types/-core crate, not duplication. This reads as undocumented, unintentional duplication that has grown into a real maintenance-risk: a bug fixed in one backend’s qlora_preflight.rs has a 720-line surface to silently miss in the other unless someone remembers to diff both crates on every change.

Files that differ meaningfully between the two backends (model.rs, inference.rs, candle_qlora_train/mod.rs, device-selection code) are expected to diverge — that’s the actual CUDA/Metal split the two crates exist for. The concern here is narrowly the 23 unmarked identical files.

Summary

Metric	Value
MENS surface total (Rust + TS)	~52,245 lines
Share of workspace Rust	~5.8%
Test-line share (Rust, lower bound)	~26%
Largest single file	tensor/preset_schema.rs, 1,368 non-blank lines
Files over the God Object Error threshold (500 lines)	26 (23 unsuppressed)
Byte-identical undocumented duplicate files, cuda↔metal	23 files, ~3,000 lines

Recommendations for whoever picks up further MENS work:

1. Triage the 23 unsuppressed Error-tier god_object findings — either they’re real decomposition debt (most look like it) or CI isn’t actually running this detector against these paths, which is its own bug.
2. Either mark the 23 unmarked identical candle-cuda/candle-metal files with vox:defactored-from (if they’re deliberately meant to diverge later and staying small enough is a real plan) or extract the large ones (qlora_preflight.rs, hf_keymap.rs, checkpoint_state.rs, manifest.rs, merge.rs) into a shared crate per the defactor policy’s own escalation rule.
3. tensor/memory_model.rs and tensor/domain_router.rs are good first decomposition targets given tensor/ already demonstrates the target granularity elsewhere in the same directory.