Add a shared round-trip corpus + fix the wire-fidelity bugs it caught (SessionStatus bitset, changeset range, origin)

[joshmouch]

What this does

Adds a shared, language-agnostic round-trip corpus under types/test-cases/round-trips/ that every client (Go, Rust, Swift, Kotlin, TypeScript) decodes into its real generated types, re-encodes, and asserts against a single canonical form — the wire-fidelity analog of the existing reducer fixtures.

Standing it up surfaced three real bugs where the generated clients had drifted from the canonical TS source, so this PR also fixes them.

⚠️ Two SemVer-major fixes (the clients were already non-conformant)

• SessionStatus is a numeric bitset on the wire (the Bitset JSDoc marker in types/channels-session/state.ts), but Rust modeled it as a closed enum and Kotlin as a signed Int — neither could hold a flag combination (InProgress | IsArchived = 72) or an unknown forward-compat bit (e.g. 2147483720). It is now a uniform 32-bit unsigned bitset: Rust u32, Kotlin UInt, Swift UInt32 (Go was already uint32).
• ChangesetOperationTarget range is range: TextRange per the spec (nested {start: {line, character}, end: {line, character}}), but the generators hardcoded a flat {start, end} integer pair that cannot represent a real range. It now references the canonical TextRange (TypeScript already did).

Non-breaking fixes

• ActionEnvelope.origin is omitted (not null) when absent, per the spec (origin: ActionOrigin | undefined; undefined = server-originated).
• Forward-compat: unknown discriminants on evolvable unions round-trip instead of throwing.

Corpus conventions

• Each fixture is one input object + an acceptableOutputs array (per @connor4312's suggestion on the previous review). Each array carries exactly one canonical entry — multiple "acceptable" outputs would risk cementing observed-but-wrong client divergence instead of pinning the single intended wire form.
• Exact-match, key-order-independent but key-presence-sensitive: null is not normalized to absent, so e.g. the origin omit-vs-null distinction is genuinely tested.
• Group A = every client agrees. Group B (017/019) = a known type carries unknown wire keys; runtime decoders drop them (acceptableOutputs[0]) and TypeScript preserves them (typescriptOutput) — a real type-system difference, documented in KNOWN-FIDELITY-GAPS.md, never a skipped assertion.

On scope

The corpus is additive; the two SemVer-major fixes are coupled to it (the corpus is what caught them, and their regression fixtures cannot pass without the fixes). I have kept them together as one coherent change, but I am very happy to split the breaking fixes into separate PRs if you would rather review or version them independently — just say the word.

[joshmouch]

Done. Every fixture is now a single input + acceptableOutputs array, uniform across the five clients.

I held acceptableOutputs to one entry on purpose: allowing several "acceptable" forms lets clients diverge instead of pinning a single wire shape, which is the whole point of the corpus. The one case where clients legitimately differ (a known type carrying unknown future keys, which runtime decoders drop and TS can't) is a group: "B" + typescriptOutput field, so that type-system difference is explicit rather than a second blessed output.

[connor4312]

thanks!