Code quality pass: precompile workload, honest size validation, leaner deps

[IanButterworth]

Note

This PR was written by Claude (Claude Code), from a code-quality review requested and supervised by @IanButterworth.

A quality pass over the whole package — no behavior changes beyond those called out below, everything sized for a minor release. One fix per commit. Verified locally: v2/v3_tiny/v4_csp/v7x still match the stored detection references at 1e-4, plus NMS unit checks, prepare_image input varieties, and all three draw_boxes paths.

Highlights

• Precompile workload — PrecompileTools was a dependency with no workload; the (recently fixed) dummy-weight path now compiles cfg parsing, model construction, and the full inference+NMS pipeline at precompile time. Time-to-first-inference: ~12.8 s → ~0.7 s (M2 Pro, warm load ~3.8 s).
• Input-size validation checks the real constraint — assertdimconform required width/height divisible by the first conv's filter count (an output-channel number, spatially meaningless; it wrongly rejected yolov7x at 416, whose first conv has 40 filters). A new max_stride(cfgvec) walks the layer blocks tracking cumulative downsample (conv/maxpool multiply, upsample divides, reorg multiplies, routes adopt their source's scale) and validates against the true maximum. Verified: 32 for all classic models, 64 for v4_p6; invalid sizes still throw; unit tests added.
• benchmark() moved behind a BenchmarkExt extension — BenchmarkTools and PrettyTables were hard deps used only by this one utility; they're now weakdeps, and calling benchmark() without them raises a MethodError with a hint. Also found and fixed along the way: the table printing was already broken under PrettyTables v3 (allowed by compat; renamed kwarg), and the hardcoded model list was stale — it's now keyed by YOLO_MODELS name and covers all 15 models at their native sizes.
• Reference integrity — get!(RES_REFS, key, computed) let a new model/image combination pass trivially by self-blessing on first run; missing references are now hard failures. The generation script (darknet-parity-asserting, per-model regeneration, same reduced test sizes as the suite) is committed at dev/generate_test_references.jl — round-trip verified.
• AllocWrappedModel parameterized — concrete field types on the default CPU entry path instead of dynamic dispatch per call.

Cleanups

• Removed verified-dead code: flipdict, createcountdict (no call sites), lhtan_grad (training-only gradient), the pre-1.9 get_extension fallback shim (unreachable on 1.10+), redundant usings, and the cu_functional guard in CUDAExt — which held a CUDA.allowscalar(false) that never executed. Notably it is deleted, not resurrected: allowscalar is a global CUDA.jl session flag that a library shouldn't impose on its users (thanks @IanButterworth for catching that in review).
• linear(x) = x replaced by Base.identity, which the codebase already used for the same purpose.
• Weights-file training metadata (seen/seen_images) is kept in cfg instead of read-and-discarded.
• Flux compat narrowed to CI-tested versions (0.14.1, 0.15, 0.16) — 0.12/0.13 were advertised but never exercised, and predate the extension mechanism.

Deliberately deferred (bigger refactors, own PRs)

• Concrete structs for the per-output Dict{Symbol,Any} post-processing state (touches Adapt/Functors traversal used by the GPU/bumper paths).
• Decomposing the ~300-line Yolo constructor into parse/assemble/precompute stages.
• The dispatch-based prepareimage.jl redesign (the #TODO: Make this multiple-dispatchy — most user-facing surface, wants its own test-backed PR).

🤖 Generated with Claude Code