Scope apply's draft write to the books an edit reaches #505

DraftRepository.apply splits the whole draft by book on every write, which takes about 3 s per apply on the large tier (#487). The plan's storage layout calls for a write to touch only the books that link a payload it changes, found through the analysis-id index that #487 adds.

This blocks 4a (#453). With the flag on, render waits for the apply, but the flag-off p90 it must match is 428 / 472 ms on large-loadable.

Done when: on the #407 tiers, an apply's time follows the books it reaches rather than the size of the draft, and npm run perf:indexes still reports no differences.