fix(op): slice accessors return copies; drop a gen group member the generator retired - #640

Two verified defects from the epic audit. Closes #532 and #572.

#532 — slice accessors aliased the graph's internal storage

Graph.Edges, Subgraph.Edges, and Subgraph.Children returned the graph's own slices, so a caller
outside pkg/op could rewrite which edges and children the graph records — leaving the construction
checksum describing contents that no longer exist. Subgraph.Edges documented the defect rather than
fixing it: "The returned slice aliases the underlying storage; callers must not mutate it directly."

All three now slices.Clone. External mutation becomes impossible rather than discouraged, and the
elements inside stay shared — exactly the sharing the rule permits. The three "must not mutate" comments
are gone because they are no longer true.

Pinned as the issue's exit asks: each accessor is written through with a deliberately wrong value and the
graph is asserted unmoved, plus a check that the units themselves are still shared so only the slice
header is defended.

#572 — a grouped-target member the generator can never emit

Every provider's grouped generation target listed gen/node_builder.gen_test.go, whose emission was
retired with NodeBuilder in phase 5. Under GNU Make 4.4.1's grouped-target (&:) semantics a member that
can never appear leaves the whole group permanently stale, so codegen re-ran on every single build.
Removed from all 16 groups, and the comment block above them corrected to describe what is actually
emitted — with the reason recorded, so the next member added is checked against the generator.

Verified rather than assumed: a second consecutive make build now runs zero codegen invocations.
Before the fix every group was stale, so it ran them all, every time.

The issue's second half — "the groups are also missing a member in the other direction" — is already
correct in the tree: all 19 providers that emit gen/module.gen_test.go list it. Only the
node_builder half was live.

What this does not fix

GOOS=linux make build still fails, and it is a separate defect, now filed as #639: GOOS/GOARCH
are ambient, so cross-compiling the product also cross-compiles the build tools — go run ./cmd/devlore-inventory produces a Linux binary and then tries to execute it on the host. The standing
cross-platform gate never caught it because make vet compiles without running anything.

Gate: make test 103 ok / 0 fail; make vet clean under darwin, windows, and linux; gofmt -l clean.