docs: say what actually guards the dynamic_message publish/take casts

[espressolee]

Comments only. No code changes — every changed line starts with //, and stripping comments
from both files before and after gives byte-identical text.

What it fixes

DynamicPublisher::publish carries this:

// SAFETY: The message type is guaranteed to match the publisher type by the type system.

It is a verbatim copy of the comment on Publisher::publish, misspelling of "explictly"
included. On Publisher<T> the sentence is true — the generic parameter ties the message to
<T as Message>::RmwMsg. On DynamicPublisher, whose whole point is that the type is a runtime
value, the compiler cannot see anything of the sort. What actually guards the cast is the
message_type comparison two lines above, and that comparison is by name
(MessageTypeName { package_name, type_name }).

The obligation neither comment mentions

Two type support libraries are involved:

	library	role
DynamicMessageMetadata	rosidl_typesupport_introspection_c	lays out message.storage
the publisher / subscription	rosidl_typesupport_c	what rcl_publish / rcl_take serialise with

So the name comparison rests on those two describing the same layout. Both constructors already
say the two libraries exist, in prose a few lines up — "This loads the introspection type
support library … However, we also need the regular type support library" — so this is known,
just missing from the place where the unsafe block is justified.

DynamicSubscription's take() is the mirror: metadata.create() lays out the storage by
introspection, rcl_take writes it through the C type support, and the comment discusses
pointer validity only.

Both are sound as written

Each constructor resolves the metadata and the C type support from the same package name in the
same function, through one ament.find_package, so they come from one install prefix and agree
by construction. This PR records that, rather than changing it. The point is that a reader
auditing these blocks is currently told the compiler is checking something it never sees.

Happy to reword or drop either hunk.

---

Generative AI disclosure

This PR was opened on 2026-08-08, before this repository adopted the OSRF AI policy in #672 on
2026-09-01. Adding the disclosure now so the record is complete:

Assisted-by: Codex:gpt-5.6-sol

AGENTS.md also asks for the trailer in the commit message body. I have left the commit as it
is rather than rewriting it, since amending would invalidate the existing CI results and any
references to the current SHA. Happy to amend and force-push if you would prefer the trailer on
the commit itself.

[espressolee]

Gentle review ping: this remains a comments-only safety-contract clarification at 15b5548b24b0db4161d46a901aab36c6324f01f0.

• Rust Stable: success
• Rust Minimal: success
• Rust Windows: success
• mergeable: true
• no runtime, ABI, or generated-code change

The issue it documents (#664) is still open; this PR only states the existing ownership/lifetime guard around the casts and does not claim to fix #664. Happy to reword or drop either comment hunk.