Backport non-breaking changes from main for Humble and Jazzy to branch v0.9

[nbbrooks]

Why

Humble and Jazzy have been stranded on 0.9.3 / 0.9.4 since early 2024 while Kilted, Lyrical and Rolling moved to 1.0.3. See #69.

main itself builds and tests clean on both. However, 1.0.0 removed public API (TypeDefinition<T>, getLockedReference()/LockedRef, details/mutex.hpp) and changed ABI. Shipping it to a released distro would break violate semantic versioning contracts. Users can source build main to opt-in to breaking changes.

The v0.9 branch takes only the additive and bug-fix subset of the 75 commits since 0.9.4, to be released as 0.9.5 — a legitimate patch bump on both distros.

Included

PR	Commits	Gives Humble + Jazzy	How it went in
#40	dd73810	Fixes UB in SerializeIntoBuffer — write through reinterpret_cast<T*> into the byte buffer	⚠️ Partial. Only the std::memcpy hunk. The rest is clang-format churn against post-1.0.0 has_TypeDefinition SFINAE overloads that don't exist here.
#60	561dcf4	inline on TypeField::operator== — fixes multiple-definition link failure	Clean cherry-pick
#62	3bb5e80	MCAPSink::finishQueueAndStop() + 3 protected DataSinkBase queue hooks	Two conflicts. The dt_tests.cpp hunk dragged in an unrelated LockedPtr test using 1.0.x-only API — dropped, kept only FinishQueue. See the restartRecording note below.
#65	c010851 14474eb de6319f 0acbf45 a355a4d 8254589	LogChannel::removeDataSink(), getNumberOfSinks(), sinks_mutex locking, logging_started race fix, new add_remove_sink_tests.cpp	Conflicts were pure clang-format drift (if ( vs if(). Took upstream semantics, kept this branch's spacing. All six needed — 8254589 fixes a race c010851 introduces.
#58	eb17fee	CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS	Clean. Verified to apply and not regress Linux; the Windows fix itself is untested here, same as on main.

Plus two commits authored for this branch, not cherry-picks — please don't try to match them against main:

• tests: drop unused id_value in FinishQueue test — this branch applies -Werror globally via add_compile_options, so it covers tests. main scoped it to target_compile_options(data_tamer PRIVATE ...), which is why the unused variable in 3bb5e80 never tripped upstream.
• fix: missing return in LockedRef move-assignment — LockedRef& operator=(LockedRef&&) declares a reference return type and falls off the end. Any consumer that move-assigns one gets -Werror=return-type, or UB without -Werror. Latent in-tree only because nothing instantiates it. This is the counterpart of Fix compile error #23 (see below).

Excluded

PR	Reason
13e10a9 "Merge branch 'new_custom'" (no PR)	Breaking. Removes TypeDefinition<T> struct API and getLockedReference()/LockedRef. This is the change that stranded these distros.
#27	Breaking. Deletes public header details/mutex.hpp, redefines Mutex as std::shared_mutex.
#24	Tests/docs/example for the post-#27 locked-ptr API. Won't compile here.
#47	ABI break (BasicType : uint8_t). Also flips BuilSchemaFromText to check_hash = false, silently weakening validation this branch currently enforces.
#40 (rest)	49b4c6c, f2dcee6 are built on the post-1.0.0 has_TypeDefinition SFINAE. Only the UB hunk extracts.
#57	README documenting the new TypeDefinition API — actively wrong on 0.9.x.
#68	ABI break — moves the constructor inline into a template. Source-compatible and it does compile on Humble, but it breaks ABI for anyone linked against the existing debs, and supersedes #17. 134 of its 190 lines are tests, which could be lifted separately if wanted.
#23	Not applicable. Patches LockedPtr::operator=, a class that only exists after the LockedRef → LockedPtr rewrite in #27 — the same rewrite that introduced #23's bug. Equivalent fix authored directly instead (above).
#37	Excluded by choice. See the restartRecording note below.
#17 #32 #36 #54 #64	Additive but out of scope for a patch release; #36 and #54 change build semantics.
#67	Not merged upstream yet.
#28 #30 #38 #49 #50 #51 #52 #53 #55	CI and release plumbing for the 1.0.x line — irrelevant to this branch, which needs its own Humble+Jazzy workflow.

One resolution worth a careful look

3bb5e80 adds a startAcceptingSnapshots() call to restartRecordingImpl(), which doesn't exist at 0.9.4 — it arrived with #37. Git auto-resolved this into the tail of restartRecording(), which is the semantically right place, but it was fuzzy auto-resolution rather than an authored decision.

If #37 is ever backported, this must be revisited or the call lands in the wrong function.

Verification

Built and tested in osrf/ros:humble-desktop and osrf/ros:jazzy-desktop:

• 23/23 tests pass on both, including the new DataTamerBasic.FinishQueue and the add_remove_sink_tests.cpp suite
• A standalone TU that move-assigns a LockedRef fails to compile before the fix and compiles after, on both distros

Net diff vs 0.9.4: +219 / −12 across 14 files. No public API removed.