fix(obs): port the source to moq_origin_request

[kixelated]

Root cause

#1772 removed moq_origin_consume from libmoq's C ABI but left cpp/obs/src/moq-source.cpp calling it. The OBS plugin has not compiled against the in-tree libmoq on dev since that PR landed.

Verified from the generated artifact rather than inferred: cargo check -p libmoq runs cbindgen, and the resulting target/include/moq.h contains moq_origin_request / moq_origin_consume_announced but no moq_origin_consume. MOQ_LOCAL defaults to the repo root, so just obs build resolves libmoq from the working tree and hits the missing symbol.

This is a dev-only regression, not pre-existing rot: moq_origin_consume is still present on main, so the plugin builds there. That is why this targets dev.

Why CI never caught it

cpp/obs is absent from check.yml and every PR/push workflow. The only job that compiles it is obs-build in libmoq.yml, and it is doubly insulated:

• it triggers only on a libmoq-v* tag push, never on a PR or branch push; and
• it builds against the published release archive (build.sh --libmoq-release "$VERSION"), not the in-tree crate.

So the break cannot surface until a libmoq-v* tag is cut from a branch containing #1772, and it would then go red in obs-build (which runs needs: release) only after that libmoq release is already live. Worth deciding separately whether cpp/obs should get a compile job on PRs; this PR does not add one.

The change

moq_origin_request is the faithful replacement. It resolves against what is announced now plus any dynamic fallback and fails when neither can serve the path, matching the removed call's announced-only, fail-fast contract. Behavior is unchanged, including the existing blank-video/teardown/reconnect response to an unresolvable broadcast.

It is not a mechanical rename: the broadcast arrives via on_broadcast instead of being returned, so the post-consume setup moves out of moq_source_start_consume into that callback. The request is just another libmoq subscription handed user_data, so it follows the contract the plugin already established for session/catalog/video:

• a lifetime reference pre-added before the request exists, released by the terminal (<= 0) callback via subscription_ref;
• the request handle stored so moq_source_disconnect_locked closes it, making the terminal fire promptly;
• user_data freed in the terminal callback.

The generation travels with the request rather than living on ctx. A reconnect can issue a new request while an older one still has a delivery in flight; a single request_gen slot on ctx would let that stale delivery read the new generation, pass the staleness check, and clobber ctx->consume. The synchronous call this replaces took expected_gen as a parameter and was immune, so keying it off ctx would have introduced a race that did not previously exist. Hence the small broadcast_request { ctx, gen } user_data.

Also folds the two duplicated failure paths into moq_source_abort_consume_locked.

Verification

• just obs check (clang-format) passes.
• The on_broadcast function-pointer type and both call signatures were type-checked against the real generated moq.h with clang++ -fsyntax-only -Wall -Wextra -Werror.
• The plugin was NOT compiled end-to-end. OBS's cmake hard-requires the Xcode generator on macOS (cmake/macos/compilerconfig.cmake), and xcodebuild on this machine fails to load IDESimulatorFoundation and asks for xcodebuild -runFirstLaunch. That is a local toolchain problem unrelated to this change, but it means the full translation unit is unverified here and should be built by a reviewer (or by a libmoq-v* tag) before merge. Draft for that reason.

Cross-Package Sync

rs/libmoq C ABI row: cpp/obs/src is this change. doc/bin/obs.md needs no update; it is a build guide and references no C API symbols (grep for moq_origin/moq_consume there is empty).

[sourcery-ai]

Sorry @kixelated, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery