Repository navigation
Expand file tree
/
Copy pathDUPLICATION.toml
More file actions
129 lines (105 loc) · 9.3 KB
/
Copy pathDUPLICATION.toml
File metadata and controls
129 lines (105 loc) · 9.3 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
# The monorepo's entries for VisualGeometryToolkitCore's `check_duplication`. Owner ruling
# (2026-11-11): every duplicate is hoisted into its one owner; an entry is allowed only
# where the two things are NOT the same concept, and its reason is the one-line proof.
# An entry that stops matching anything is itself a finding: remove it. Paths are
# relative to this file.
# API note (owner ruling 2026-11-11, D-mad): VisualGeometryToolkitCore._mad_scale_norms(; norms, codimension)
# is a STACK-INTERNAL entry point, not a duplicate: the one median-of-norms scale that the public
# `mad_scale(residuals)` itself calls, reached (qualified) by VisualGeometryToolkitEstimation only where a
# kernel holds just the norms (fit_rotation's angles, the radial-homography lift seed). There is no
# public norms form, so a reader of `mad_scale` never wonders whether norms are re-centred.
# Declared, not an entry (an entry the checker does not flag is itself a `:stale_exception`): VisualGeometryToolkitEval's
# src/code_identity.jl hashes with SHA.sha256 directly (`file_sha256`, `tree_sha256`, `_toml_hash`), not through Core's
# `sha256_hex`, because src/cli.jl includes it, and a project's runner launcher includes src/cli.jl by path from a bare
# `julia` before the project's environment (and so VisualGeometryToolkitCore) exists; the digests are the same SHA-256.
# Test-input fixtures (owner ruling 2026-11-13: they go here). Restored in wave 113: 12ab92d0 removed them because
# the gate, which reads the stack repositories, did not match them while CI did; the idiom cap now counts this
# repository's definitions only, so both find them.
[[similar]]
functions = ["VisualGeometryToolkitCore/test/data/chain_equivalence/core_values.jl::canonical",
"VisualGeometryToolkitCore/test/fixtures/fixture_literals.jl::lit"]
reason = "different output grammars: `canonical` writes the chain-equivalence record's canonical text (\",\" with no space, sorted keys, every number as a Float64), `lit` writes Julia source literals (\", \", a Float32 kept); no input gives both the same text"
[[similar]]
functions = ["VisualGeometryToolkitCore/test/fixtures/fixture_literals.jl::rows",
"VisualGeometryToolkitIO/test/migrate_versions_tests.jl::vec_"]
reason = "different output grammars: `rows` writes a multi-line Julia array literal of literals, `vec_` a one-line JSON array of numbers inside a hand-written record"
[[name]]
module = "VisualGeometryToolkitEval"
name = "_TOML_DEFAULTS"
reason = "not code but each package's own data: Eval's parsed configs/default.toml and Estimation's are different files with different tables (experiment defaults vs RANSAC defaults); both are parsed and read through Core's one `parse_defaults`/`toml_default`"
[[name]]
module = "VisualGeometryToolkitEval"
name = "Pools"
reason = "different contents: Eval.Pools reads and writes pool FILES (pool_io.jl); Core.Pools defines the pool RECORD types (PoolPair) Eval.Pools builds; no function of one restates the other"
[[backend]]
file = "VisualGeometryToolkitCore/ext/VisualGeometryToolkitCoreCUDAExt.jl"
function = "__init__"
reason = "not a CPU/GPU twin of any code: the CUDA extension registers the device queries (free memory, reclaim) a KernelAbstractions backend cannot answer, and answers them for its own backend only; the policy's allowed device-query helpers"
[[backend]]
file = "VisualGeometryToolkitCore/src/geometry/transforms/pointwise.jl"
function = "p::OnBackend"
reason = "not a CPU/GPU twin of any computation: the stack's one kernel-launch policy (the owner's ruling, 2026-11-12) instantiates the SAME kernel on every backend and chooses only its workgroup per backend — a one-item kernel runs one item per workgroup on the CPU (every thread busy) and one warp per workgroup on a device (one item there would idle 31 of 32 lanes)"
[[backend]]
file = "VisualGeometryToolkitFeatures/src/asift/driver.jl"
function = "_chunk_warps"
reason = "not a CPU/GPU twin of any computation: it sizes how many warps stay resident (an allocation), and only the free memory differs by backend — the host's on the CPU, the device's less the room the ONNX Runtime CUDA sessions' arenas may still take on a GPU (those arenas are device memory; the CPU's warps share no device with them); the same kernels run on every backend"
[[backend]]
file = "VisualGeometryToolkitFeatures/src/onnx_runtime.jl"
function = "_onnx_device_free"
reason = "not a CPU/GPU twin of any computation: it reads the free device memory around an ONNX Runtime CUDA session's creation and runs, to measure what that session's arena took; on the host there is no device memory to read (an arena is device memory), so it reads nothing"
[[backend]]
file = "VisualGeometryToolkitCore/src/quality/gpu_tier.jl"
function = "gpu_tier"
reason = "not a CPU/GPU twin of any computation: the test runner's one decision whether a GPU tier runs at all reads which backend the default device is (the backend require_gpu hands the GPU items); no kernel or arithmetic is chosen by it"
[[backend]]
file = "VisualGeometryToolkitCore/src/quality/gpu_tier.jl"
function = "gpu_problem"
reason = "not a CPU/GPU twin of any computation: a GPU-tier worker's re-check that its default device is still an accelerator, so the GPU items error rather than run on the CPU path"
# check_splat's exceptions (a splat of a collection whose length is data, VisualGeometryToolkitCore's `check_splat`):
# each is a splat whose length is fixed by a type or bounded by construction, or an API with no collection form.
[[splat]]
file = "VisualGeometryToolkitEval/src/Experiments.jl"
function = "run_isolated"
reason = "Base's withenv takes its pairs only as varargs (no collection form); the count is the few variables one isolated driver run sets, so withenv is compiled once per such count, never per datum"
[[splat]]
file = "VisualGeometryToolkitEval/src/tuning/toml.jl"
function = "_read_struct"
reason = "fixed arity: the values are the fields of the struct type `C` read from TOML, one per field, built positionally by `C`'s own constructor; their count is fieldcount(C), one per type"
[[splat]]
file = "VisualGeometryToolkitCore/src/base/fit_result.jl"
function = "propertynames"
reason = "the names are a function of the type `FitResult{P, M}` alone (its metadata's fields and the nested ones), so the tuple has one length per FitResult type, as the type's own fieldnames do; Base's propertynames contract is that tuple"
[[splat]]
file = "VisualGeometryToolkitCore/src/robust/normaliser.jl"
function = "truncated_log_normaliser"
reason = "QuadGK takes a split domain's breakpoints only as varargs (2.11 has no vector form); their count is ⌊log₄(r_max/c)⌋ + 3 ≤ 63, a few distinct values over the widths in use, so `quadgk` is compiled a few times, never per datum"
[[splat]]
file = "VisualGeometryToolkitCore/ext/VisualGeometryToolkitCoreTestExt/contracts.jl"
function = "test_scoring_contract"
reason = "not this code's splat: Test's `@inferred` of a call with keywords expands to its arguments gathered in an `Any[…]` and splatted back, so the length is the call's written arity (fixed); the expansion carries the call's own line, so check_splat cannot tell it from a written one"
[[splat]]
file = "VisualGeometryToolkitEstimation/src/fitting/division_camera_step.jl"
function = "CameraModel"
reason = "`Division(k₁, …, k_Q)` takes the lens's coefficients as varargs (Core's one constructor) and Q is the division model's order, the fit's configured number of coefficients (1–3), fixed for a run, never a data length: `Division` is compiled once per order"
[[splat]]
file = "VisualGeometryToolkitEstimation/src/config_toml.jl"
function = "_tagged_from_toml"
reason = "a tagged TOML table builds its type `T` from its fields' values in field order: the arity is `fieldcount(T)`, fixed by the type, so `T` is compiled once per tag type"
[[splat]]
file = "VisualGeometryToolkitIO/src/migrate.jl"
function = "write_staged"
reason = "the public API's callback receives one staged path per output it names, `f(staged...)`: its arity is the caller's written number of outputs (1 or 2 at every call site, a fixed arity per call), and the outputs are varargs Pairs by design"
[[splat]]
file = "VisualGeometryToolkitIO/src/rig_schema.jl"
function = "_rig_from_record"
reason = "VisualGeometryToolkitCore's `Rig` holds its sensors as a Tuple (each sensor's type in the rig's type), so a rig read from a file is built from a tuple of its sensors: one length per rig layout, the length the type itself carries"
# The sensor model's one black subtraction (owner rulings D1–D13, 2026-11-13): the LibRaw decoder hands LibRaw its own
# per-channel black for the display demosaic; it subtracts nothing in the stack.
[[pattern]]
use = "black_subtraction"
file = "VisualGeometryToolkitFeatures/ext/VisualGeometryToolkitFeaturesLibRawExt.jl"
reason = "not a black subtraction: the decoder writes the mosaic's per-channel black back into LibRaw's `cblack` for LibRaw's own display demosaic (the :demosaic image), less the common black LibRaw keeps apart; no stack value is linearised there"
[[pattern]]
use = "sensor_component"
file = "VisualGeometryToolkitIO/src/sensor_characterisation_schema.jl"
reason = "the vgio/sensor_characterisation kind's codec is the ONE reader that turns a file into sensor-model components (ChannelBlack, Dark, Darks, ReadNoise, PoissonGaussian, RelativeNoise, DefectList) and a SensorCharacterisation, through their public constructors"