docs: the merged image's cost, measured — and it corrects my own estimate (4296)

I wrote in ffcd130's Dockerfile comment that merging ML in means "everyone
pulls it, including the many who will never turn tagging on", framed as a
real cost the milestone accepted. My working estimate behind that was ~4GB.

Measured from run 7273's build-web log:

  torch 2.12.1+cpu wheel      192.3 MB
  torchvision 0.27.1+cpu        1.8 MB
  transformers / onnxruntime / opencv / sklearn and friends
                               62.0, 35.3, 23.6, 16.7, 12.3, 9.2, 6.9 MB
  largest newly-pushed layer  222.07 MB

The ML code adds a few HUNDRED MB, not gigabytes. The `--index-url` CPU
resolution is what makes that true — the default PyPI torch wheel carries the
CUDA runtime and is ~2GB by itself, and the log confirms 2.12.1+cpu resolved,
so it is working as intended rather than as intended-but-unverified.

Why this matters beyond a comment being wrong: it settles the trade this step
was explicitly asked to weigh and could not, and it reverses how close the
call looked. Baking the weights in adds ~3.5GB to every pull for a feature
many adopters never enable; shipping the code and fetching on demand adds
~350MB. An order of magnitude, where the estimate had them within 15% of each
other. Off-by-default is not a judgement call here, it is arithmetic.

The gigabytes were always in the MODEL, and the model is not in the image.

Two things NOT measured, still: the total image size (the push only transfers
layers the registry lacks, so a push log cannot give it) and the per-slot
resident RAM, which stays flagged `measured=False` in the lane table and
renders as "about" in the UI.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LVjrnpQjRgHdvq95rASoiR

Dockerfile

@@ -51,19 +51,35 @@ RUN pip install -r requirements.txt
 # in one process tree, a second image would mean the `ml` lane could never be
 # enabled from the UI — there would be no worker in this container to enable.
 #
-# The COST, stated because it is real and falls on every adopter: this adds
+# THE COST, MEASURED from run 7273 rather than guessed — and it is far
-# torch, torchvision, transformers, onnxruntime and opencv to an image that
+# smaller than the estimate this comment first carried, which said "everyone
-# previously carried none of them. Everyone pulls it, including the many who
+# pulls ~4GB":
-# will never turn tagging on. That is the trade the milestone accepted for
-# being able to offer the lane as a switch rather than a second deployment.
-# What it buys back is that nothing downloads a MODEL until the switch is
-# thrown — the weights are not baked in, and rule 164 permits that only
-# because the feature is optional and clearly off.
 #
-# CPU-only torch from the PyTorch CPU index. The default PyPI wheel bundles
+#   torch 2.12.1+cpu wheel     192.3 MB
-# the NVIDIA CUDA runtime (~5.6GB of layer) and nothing here uses a GPU — the
+#   torchvision 0.27.1+cpu       1.8 MB
-# GPU agent is a separate service with its own image. `--index-url`, not
+#   transformers / onnxruntime / opencv / sklearn and friends
-# `--extra-index-url`: the latter would let pip resolve a +cu wheel anyway.
+#                               62.0, 35.3, 23.6, 16.7, 12.3, 9.2, 6.9 MB
+#   largest newly-pushed layer 222.07 MB
+#
+# So the ML code adds a few hundred MB to the pull, not gigabytes. The CPU
+# index is what makes that true: the default PyPI torch wheel bundles the
+# NVIDIA CUDA runtime and is ~2GB on its own.
+#
+# The GIGABYTES are in the MODEL — ~3.5GB of SigLIP weights — and those are
+# NOT in this image. They arrive only when the operator enables the lane,
+# which is what lets rule 164 permit a runtime fetch at all ("optional and
+# clearly off"). That also settles the trade this step was asked to weigh:
+# baking the weights in would add ~3.5GB to every pull for a feature many
+# adopters never enable, against ~350MB for the code that makes the switch
+# available. Off-by-default wins by an order of magnitude, which was NOT
+# obvious before measuring — the estimate had the two costs within 15% of
+# each other.
+#
+# `--index-url`, not `--extra-index-url`: the latter would let pip resolve a
+# +cu wheel anyway, and the whole saving above depends on it not doing that.
+#
+# CPU-only torch from the PyTorch CPU index. Nothing here uses a GPU — the
+# GPU agent is a separate service with its own image.
 RUN pip install --index-url https://download.pytorch.org/whl/cpu \
         "torch>=2.12,<3.0" "torchvision>=0.27,<0.28"
 RUN pip install -r requirements-ml.txt