refactor(libmoq)!: replace the client setters with a config struct

[kixelated]

Summary

• libmoq ships as a staticlib only: the release tarball packages include/moq.h beside lib/libmoq.a, nix/overlay.nix installs just the archive, and crate-type = ["staticlib"] builds no cdylib. A caller therefore always compiles against the header matching the archive it links, so appending a field to a C struct is a recompile, not an ABI break.
• The moq_client_set_* / moq_client_get_* functions existed to avoid that break, and the docs said so. Since the break can't happen they bought nothing but surface, so they collapse into one moq_client_config struct plus moq_client_defaults(), which replaces every getter at once.
• moq_client_connect folds into moq_session_connect, whose new config parameter is NULL for the defaults. One connect function instead of two.
• Exported C functions: 121 → 81.

Knobs whose default is not zero carry a has_* flag rather than being read directly. websocket.enabled defaults to true and the reconnect backoff to 1s/2/5s/10s, so a caller who zeroed the struct and set nothing would otherwise have silently disabled them. The flag is also what separates "leave it alone" from an explicit value: backoff_timeout_ms = 0 with the flag set means retry forever.

moq_client_defaults returns by value. With no out-pointer there's no handle to look up and nothing to reject, so the int32_t status channel had no failure left to report.

Also in this change:

• moq_video_config's coded_width/coded_height stop being *const u32. Zero already means absent for a dimension, so each keeps its own presence and a half-specified catalog still round-trips (collapsing them behind one flag invented a zero for the missing half).
• build.rs sets pragma_once. cbindgen.toml is never loaded, so the generated header had no include guard — fine until moq-settings.h started reaching it down a second include path.
• moq_publish_{video,audio}_config no longer claim that re-declaring a rendition name replaces it. hang's insert returns Duplicate; pre-existing doc bug.
• The docs drop their claims that a dynamic library ships (libmoq.so / libmoq.dylib / moq.dll never existed).

Public API changes

Breaking, hence dev — all in rs/libmoq's C ABI:

• Removed: moq_client_create, moq_client_close, moq_client_connect, and every moq_client_set_* / moq_client_get_*.
• Added: moq_client_config (struct), moq_client_defaults() (returns it by value).
• Signature changed: moq_session_connect gains const moq_client_config *config after url_len.
• Field types changed: moq_video_config.coded_width / .coded_height are uint32_t (were const uint32_t *).
• Header: moq.h now carries #pragma once.

Rust-side libmoq::Error::ClientNotFound is gone (the handle it named no longer exists). libmoq is a staticlib, so its Rust surface has no consumers.

Test plan

• cargo nextest run -p libmoq — 64/64 pass, including a new a_half_specified_coded_size_round_trips that forwards a width-only catalog between broadcasts, and a_zeroed_config_is_the_defaults / defaults_report_what_a_zeroed_config_dials pinned against client::Config::default() so retuning a default without following through to C fails here.
• cargo clippy -p libmoq --all-targets and cargo fmt — clean.
• cargo doc -p libmoq --no-deps — clean (several intra-doc links pointed at the removed functions).
• A standalone TU compiles moq.h twice and exercises every shape cpp/obs depends on: the 7-arg dial with NULL, scalar dimensions, and the moq_client_config fields the settings builder assigns.

Cross-Package Sync

rs/libmoq C ABI → cpp/obs/src and doc/lib/c are both updated. cpp/obs/src/moq-settings.cpp drops ~130 lines of setter calls for field assignments; MoQSettings::Config now owns the strings the config borrows, since libmoq reads them during the dial rather than copying at set time.

Reviewer notes

• cpp/obs is not compiled here. PR CI never builds it and there's no libobs on this machine, so just obs build / just obs test are still the real gate. An adversarial review pass caught that three OBS call sites hadn't been migrated at all; those are fixed, but the plugin has only been checked against the header, not built. feat(obs): compile the plugin in CI, and type-check it without obs-deps #2867 would close this gap.
• Per-field validation moved to dial time. A bad fingerprint used to fail at its setter; it now fails at moq_session_connect with the reason in moq_error(). BuildConfig can no longer name the offending field — it still refuses to start, just less specifically. That's the one thing the setters genuinely bought.
• Overlaps heavily with feat!(libmoq): add an optional track name to moq_publish_media #2831, which touches the same api.rs / test.rs / OBS files. Whichever lands second will need a rebase.

(written by Opus 5)

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b3fa6e4572

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".