Store full Source alongside legacy source_identifier on all sourced commands

[ryantaylor]

Summary

• BuildSquad/BuildGlobalUpgrade (SourcedPbgid), UseAbility (SourcedAbility), and CancelProduction (SourcedIndex) previously exposed only a legacy truncated u16 source_identifier — the fully parsed Source was thrown away right after wire parsing (Source::parse always runs and decodes it in full, regardless of command type; the truncation was purely a backward-compat holdover from before the Source enum existed).
• Every other sourced command variant (Targeted, SourcePbgid, Sourced, Ability) already exposes the full Source. This closes that gap.
• Each of the three structs now stores both source_identifier: u16 (kept for existing consumers) and source: Source (for new code — Self::source is now the recommended accessor), rather than replacing one with the other, per the precedent already set by a2eb90a (CMD_CancelConstruction).
• CommandPayload::source() now returns Some(..) for BuildSquad/BuildGlobalUpgrade/UseAbility/CancelProduction, where it previously always returned None.
• SourcedPbgid/SourcedIndex/SourcedAbility are no longer Copy (only Clone) since Source::Squads holds a Vec<u32>, consistent with every other full-Source struct in the crate.
• Updated the cohlib-rb Ruby marshalling to include source in the hash for these three payload shapes.
• Added a regression test (legacy_sourced_variants_now_expose_full_source) guarding that these four variants now carry a full source across the fixture corpus.

Test plan

• cargo test --workspace — all green
• cargo clippy --workspace -- -D warnings — clean
• cargo fmt --check — clean
• cargo build -p cohlib -p cohlib-rb --features cohlib/magnus — clean (Ruby-facing binding compiles)
• cargo run --release --example vault_regression against the ~9.3k-replay real-world corpus — 0 panics, 0 parse errors, 0 unexpected Unknown fallbacks (confirms no command in the corpus hits the Source::Squads panic path for these four newly-exposed types)

🤖 Generated with Claude Code