docs(rooms): a todo list, in the shape this repo already uses for them

[stormer78]

tasks/ already holds vta-architecture-todo.md and vtc-architecture-todo.md — same status legend, same sizes, PR numbers recorded as things merge. Rooms had none, so its backlog lived in conversation. That is exactly how data-rooms-epoch-anchoring.md came to sit marked "open question, not a plan" for two days on a question nobody outside one thread could see.

Five phases, in dependency order:

Phase 0	Verified reads — shipped, with PR numbers, because each is the answer to when did this stop being true
Phase 1	Read through the agent — R1.1–R1.5
Phase 2	The witnessed anchor — A2.1–A2.4
Phase 3	The surfaces — plugin client, then the console panes
Phase 4	Off the arc — self_update, commit relay, js-yaml, prune, ZK, Merkle caching, join ceremony

It opens with the three decisions that are settled and not to be re-litigated, since two of them were asked twice in two vocabularies before being answered once.

What is deliberately written down rather than left to be re-derived

• R1.3 is keyed by room, not host, with the reason. That is the line that buys a fourth comparison the specification does not list — two hosts of one room disagreeing at one headVersion — and it is one word away from being wrong.
• R1.4 depends on R1.3, explicitly. A refusal that lasts only as long as the process that noticed is not a refusal.
• R1.5 is "the words", sized and phased like any other task, because serve reads, refuse writes is a rule nobody would guess and a member who is shown only a failure concludes their own agent is broken.
• X4.1 states what RoomGroup::self_update()'s absence actually means: "a room renews" currently means "a room adds or removes somebody", which is not what renewal means — and §9's lifecycle and every anchor ride renewals.
• X4.3 records that the standing dependabot "high" is dev-only, via @redocly/openapi-core, and that the real cost is a permanent alert that trains everyone to ignore the list. So nobody repeats the investigation.
• X4.6 keeps the reason the Merkle store is deliberately uncached: a stale root is worse than a slow one, because the failure it produces is an honest host appearing to equivocate.
• X4.5 is [!] blocked rather than [ ] not-started — it waits on a working group, not on us.

Documentation only.

[affinidi-appsecurity-bot]

🛡️ AI Agentic Security Code Review — all clear. We checked this change and found nothing to report. Keep shipping secure code!

Note: for major, breaking, or feature-introducing changes, you can always request an in-depth review from the security team.