feat: migrate the backlog into the per-item store, and gate it against the table (Q889)

[karlkfi]

Phase 2 of Q889: the backlog becomes 173 files under docs/queue/, one per item, so priority is the rank key rather than a line's position in a shared table. Two sessions taking the top two items now touch two different files.

Follows #1595, which landed phase 1's tooling.

What changed

62 titles were rewritten by hand. The store caps a title at 72 characters and the old table capped only the Notes cell, so 62 of 173 were over, the longest at 130. This was the work; the migration itself was one command. Rewriting is judgement rather than truncation, so each was done individually: where a test identifier was the item's only handle it stayed (Q549 sits at exactly the cap), and where the plan doc or the trigger already named it, it moved there.

The migration is reconciled, order included. The ID sets match both ways, the Queue's 106 rows are the store's first 106 in order, and the remaining 67 are Deferred then Flake watch in order. Order is the claim worth checking, because the table's order is the priority and a count cannot see it scrambled. All 29 flake-watch rows arrive deferred carrying flake, which is the plan's decision 4 satisfied by migrate itself, with no extra work.

queue-drift-check is new, and it is why this could land separately at all. The plan made a drift check the condition for splitting phases 1 and 2 out of the atomic switch. Until phase 3 moves the consumers, the backlog is written down twice and an edit to one side alone is silent in the direction that matters. The check re-runs migrate into a throwaway store and compares the loaded items, so what is compared is queue.py's own reading of the table rather than a second parser free to drift from it. It retires itself once docs/STATUS.md is deleted in phase 6.

Two things that were nearly missed

mdreflow folded the store README's Conventions block into one paragraph, and rule 11 anchors the label vocabulary to a line starting **Labels:**. STATUS.md survives the same treatment only because its equivalent block carries trailing two-space hard breaks. The checker exiting 2 on an unreadable vocabulary is the sole reason this was caught: had it treated that as nothing to check, rule 11 would have passed green while enforcing nothing, for as long as the store exists.

The rules suite only ever filed inline labels: [a, b], which is the one form migrate never writes — every real item carries a YAML block list. The checker parses both, so nothing was broken, but nothing held it to that either. Found by a probe of my own using an inline-only parser, visible only because the reconciliation printed both directions rather than the empty one. A one-sided read would have said "no undeclared labels" and been believed.

Testing

• make check green.
• queue.py lint: 173 item(s) OK. check-queue-rules.sh: 173 items, rules 8, 9 and 11.
• check-queue-drift-test.sh: 12 checks, every field in both directions, plus the two controls they pull against (a rewrapped note is not drift; a README is not an item).
• The new block-form label case was falsified by deleting the mechanism: with the block branch removed it goes red and both inline cases stay green, so it is genuinely the only coverage for the shape the store is written in.
• The reflow's round-trip claim was verified by rendering the whole store before and after and diffing, with a control asserting the diff could see a change, rather than by reading read_item.

What is not in this PR

Phases 3 to 6: switching the 53 consumers, rewriting 218 anchors, the /dev/queue/ render, and deleting docs/STATUS.md. The drift gate holds the interim.