refactor(renderer): separate prop lifetimes and value writers

[smiggleworth]

The host-prop coordinator mixes diff ordering, DOM snapshots, bindings, handlers and refs. Separate orchestration, transactional writers, lifetime ownership and controlled-select settlement into private modules; scalar dispatch uses private class/style/form/HTML writers. All 49 original function bodies are unchanged, including their compiled production DOM-bundle bodies. Existing internal entrypoints and public package contracts stay intact.

The mixed lifetime tests exercise fresh and adopted hosts: a later writer fails after listener/binding/ref replacement is prepared, the old listener and binding remain active, replacement refs do not attach, retry preserves element identity, and replacement/release clean up once. A failed initial adoption separately proves provisional bindings/listeners are discarded and retry installs each lifetime once. All three cases pass on both the original and extracted implementations.

Validation:

• Before/after focused characterization: 489 renderer/DOM/hydration cases pass. The architecture seam assertion fails before extraction and passes after it; guards exercise static/re-export/dynamic imports and preserve the acyclic core graph.
• Complete Node 24.21.0 npm run check: 433 unit, 92 repository, 2,693 jsdom, 158 browser and 12 installed type cases; formatting, lint, build, public types, publint and package artifact checks pass.
• Three alternating captures on the node-refactor base and this candidate, all invariants and positive samples: 60-field form hydration medians 3.2809→3.2460 ms; 2,000 keyed rows through a component boundary 58.2465→62.6177 ms (+7.5%). The source-module result is retained, not dismissed or rebaselined.
• Follow-up production-artifact diagnostic uses the built public state/boot/control/JSX/testing entrypoints, asserts complete forward/backward ordering and retained row identity, then times the same fixture on both artifacts. Three alternating captures: baseline 37.3167/39.8709/33.3573 ms, candidate 37.5377/35.5794/38.7362 ms (median +0.6%, 101–122 samples). All 49 compiled prop function bodies and their local call shapes are identical. This supports a source-runner/module-boundary explanation for the first workload's delta; it is not a general speedup or absence-of-regression claim.
• Ownership documentation describes initial/adopted/committed writes and post-settlement ref errors. Readiness records merged refactor(renderer): split node dispatch from hydration and lifecycle coordination #738 with its successful dispatched merge-head CI and the still-open CLI performance gate. No versions, tags or publications change.

Closes #739.