The AOT lane stops being a pure consumer: a cell-discovery miss now chooses a
recipe and mints out of process, every decline names itself on the wire, and
aot_mint_phases starts carrying rows. Also: the boot-span ladder runs on the
real path (pgw#797), AOT host compile is 7-9% cheaper (pgw#793), cell entries
compile K-wide (pgw#809), cell receipts are algorithm-agnostic (th#1303/#807),
the conversion producer publishes v2 (th#1303 phase 3), and cell discovery
names an authz refusal (th#1335).

Two defects the release gate itself caught and fixed:
* chaos was not mypy-clean (22 errors; 20 predate v0.79.0), one of them a live
  AttributeError on every degree>1 boot.
* a background warm/heal compiled under a different intra-op thread count than
  the thread it healed for, so dynamo's GLOBAL_STATE guard made the entry
  unservable and heals burned budget toward permanent volatile.

CI 30596169781 green on this exact tree: 2184 passed, 22 skipped, 1 xfailed.