Queue phrase alignment after the last kind-word round, and let it wait while one is pending

[WaylandYang]

First of the alignment-cost cuts after #918/#920.

Kind-word alignment queued phrase alignment at the end of every round, including a round that had just queued another round of itself because new words had arrived. On the typed-graph bench (run 3) the second kind-word round (562 words) and the first phrase round (1010 signatures) ran side by side; the kind words re-bound classes under the signatures, their fingerprints changed, and the next phrase round re-decided 713 of them. Two fifths of the phrase decisions in that run were repeats.

• Kind-word alignment queues phrase alignment only from a round that queues no further round of itself.
• Phrase alignment, whichever entry queued it, checks for a queued or running kind-word job on its base before taking its lock; if there is one it queues itself for half a minute later (enqueue_unless_queued_after, at most one queued) and returns.
• jobs::pending_for_kb looks a job kind up by the base in its payload, since a reask payload carries an extra key.

No decision changes; only when the phrase rounds run. Verified: the kind-word reask test asserts phrase alignment is not queued while a reask is pending and is queued from the final round; a new phrase test asserts nothing is asked while a kind-word job is queued, one deferred round exists, and the votes are asked once it is gone. Alignment suites 14/14 and 11/11, workspace clippy clean. A bench group on this branch follows in the README before merging.

🤖 Generated with Claude Code

[WaylandYang]

Measured before merging (README, 第五次真跑): two groups on this branch and one dev control at the same hour. Signatures are clean (about 1000 transitional none rows gone; phrase decisions 1847 → 980), recall unchanged, tokens nearly unchanged because the repeated decisions were cheap. Judged precision with 600 facts: 77.2% / 81.8% on the branch against 81.7% / 82.0% on dev, indistinguishable; the 200-sample figures swung by ten points on the same base, so the bench's default is now 600.