Skiff to full: execute the in-tree plan, phases 1-5

[sshaplygin]

Part of #63.

The plan is in-tree and authoritative: docs/skiff-full-support-plan.md. This issue tracks its execution and nothing more — anything that reads differently there wins.

It came out of auditing this crate against the real C++ implementation rather than against its own idea of Skiff, and the shape of the finding is worth repeating: the bytes are right, the vocabulary is short. Nothing in phases 1–3 changes what this crate writes today; every item is a type, a rule or a framing tag that is currently absent.

• Phase 1 — serialization parity. Byte-level, small, and every later phase encodes through it, which is why it is first.
• Phase 2 — table schema matching. Where a cluster-written schema is most likely to be refused today: a control column in the wrong place is a real schema someone will hand us, not a hypothetical.
• Phase 3 — YT logical types. The largest, and it depends on both above. decimal and list<T> cannot be expressed at all right now.
• Phase 4 — typed rows. Ergonomics; can slip without blocking anything.
• Phase 5 — a standing gate against C++. What keeps the other four from decaying.

Demoted below dynamic tables by the decision of 16 August 2026. That is about order, not about doubt.