perf(octane): render computes, commit writes — runtime size redesign (#2043) - #2044

Every native read source carried a DevTools inspect closure, and the
native read driver and retry leases carried inspectScope/inspect, so every
production signal bundle shipped inspectNativeNode and
native-read-inspection.ts. DevTools only install the native read inspector
under __OCTANE_PROFILE_ENABLED__, so these members now exist only there.

DOM bindings asked inspect() whether an observed source's owner had
retired. That is functional, so sources gain a cheap retired() member;
graph, view, historical and transition-candidate sources implement it.

shell_ app (oxc), full: -1,106 raw / -302 gzip.

Existing tests pin both sides: behavior-root's "ends quietly when a scope
observed only through a ... retires" cases (dev and prod) fail 8/8 when
the binding check keeps using inspect(), and the octane-signals-profile
devtools tests fail 3/3 when graph sources lose inspect.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

React's server never runs an error boundary. ssrTry no longer renders a
@catch arm or an ErrorBoundary fallback: an error, or a rejected use(),
reports through onError and leaves the nearest Suspense boundary's
fallback, marked for a client render, in renderToString, prerender, and
streams before and after the shell. Buffered @try/@pending boundaries now
recover like JSX Suspense, with a new oct-client-error marker. A boundary
without @pending passes the error on, so with no Suspense boundary above
it the render fails.

The server no longer serializes rejection seeds, records isolated seed
runs, or compiles catch arms. The client drops the server-catch adoption
and replay path (takeServerCatch, replayCaughtTry, settleServerCatch,
adoptServerCatch, fallBackServerCatch, rejection seed decoding) and the
catch-arm adoption parameters of switchToCatch. A boundary the server
marked as errored, buffered or by $OCTRX, renders afresh and reports
error 344 through onRecoverableError, unless its catch arm takes a throw
in that render. Error codes 23 and 45 are retired.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs/ssr.md gains a Render errors section and drops the rejection seed
and server catch arm text; differences-from-react.md replaces the server
catch paragraph. The app-core root boundary comment and the React compat
pages follow, and a minor changeset describes the migration for apps
that relied on a server-rendered catch arm.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Every graph node stored its state as two objects (a NodeState wrapping a
snapshot), so each set() on a writable signal, and each recomputation of a
derived value, allocated both, through a conditional spread, inside a
signalBatch closure.

A node whose state is plain and ready is now value-only: `last` holds the
value and the ready state, which is also its retained state, is built only
when something asks for `state` or `lastState` (snapshot(), a public
subscriber, serialization, a transition fork). Writable cells start that
way; plain derived$ and compiler-selected scalar computations return their
value through evaluate() and stay that way while their dependencies carry
no activity and they read no foreign owner. Pending, failed, async and
cross-owner results keep stored states exactly as before, and the
accessors make every existing reader, including the transition fork
machinery, see the states it saw before. get() reads a value-only node
without building a state; candidate, historical and adoption reads still
go through readNode. set() no longer allocates a batch closure.

Resource cells installed their retry as an own property after
construction, which gave async nodes their own hidden class. The retry now
lives in a declared retryRequest field: %HaveSameMap reports one map for
writable, derived, failed and async nodes (async differed before).

Measured on a production esbuild bundle of octane/signals (ns/op, median
of 7, same session, scratch harness):
  set, no subscribers        44.8 -> 14.1
  set + get                  65.7 -> 28.2
  set + native subscriber   122.0 -> 77.4
  set + public subscriber   123.8 -> 108.8
  set -> derived -> native  189.7 -> 119.2
  batch 10 sets/10 derived 2218.6 -> 1702.9
scoped-signals bench (--sizes=100,1000, 4 interleaved rounds, median),
candidate/base geomean: write_sparse 0.880, batch_write_all 0.895,
create 0.898, write_equal 0.950, switch_dependency 0.953,
read_cached 1.030, dispose 1.038, continuous cycle 1.103.

shell_ app (oxc), full: +1,314 raw / +346 gzip. The fork machinery keeps
every async-state helper reachable, so no code drops out yet.

The new signals-state test pins equality cut-off for derived values: it
fails with [1, 1, NaN, NaN, 0, 0, 0] when a recomputed equal value is
treated as a change.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Each graph node built a native read source per read mode as an object of
five closures, and a declaration view built a second graph source plus a
forwarding object of five more. The field mapping that caches a source
per read mode was copied into readNode, unchangedLiveRead, the view
source and transition-candidate, which reached the graph through the
candidate bridge's createNativeSource.

A source is now one GraphReadSource (node, read mode, optional view
record) with prototype methods; a view's source forwards to its canonical
cell once promoted, exactly as the closures did. DevTools inspect is
installed on the prototype only under the profiling flag.
ScopedNode.nativeReadSource(read) owns the per-mode cache, so
transition-candidate calls it on the node it already holds and the bridge
loses createNativeSource.

The other 21 bridge entries stay: transition-action, transition-candidate
and candidate-producers sit in the renderer's module graph (runtime.ts
imports transition-candidate, which imports transition-action; islands
builds import candidate-producers), and bundle-boundaries asserts that an
ordinary client never resolves signals/graph.ts. The symbol-tagged notify
functions stay too: transition-action discovers presentation consumers
through them, and the dom-binding modules publish them.

shell_ app (oxc), full: +104 raw / -3 gzip (a closure-based variant of
the same centralization measured -48 / -33, at five closures per source).

The new pending-promotion test reads a declaration view natively, promotes
it, and checks that the render's witness stays valid and its lease follows
the canonical cell; it fails with "expected NaN to be 1" when a promoted
view source stops forwarding.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…ptors

Element descriptors had their own prop writers (applyDeoptProps,
patchDeoptProps, applyDeoptProp, setDeoptClass) beside setSpread, with the
same identity diff and removal path. Descriptor hosts now call setHostProps,
the body setSpread already used; setSpread keeps only the spread-only work
(children tracking for raw HTML, skipped keys, the function form action) so
descriptor-only bundles do not retain it. A DESCRIPTOR_HOST flag keeps the
descriptor-specific behavior: unread scoped children, ref stamping for raw
teardown, React's input checkedness initialization, and identity-kept raw
HTML. hostComponent keeps applyHostProps, whose live re-assertion contract
differs on purpose.

shell_ app (oxc): full gz 122,309 -> 122,291 (raw +188), base gz 58,005 ->
58,047 (raw +176); bundles with both spreads and descriptors drop the
duplicate writers. js-framework octane-ts (plain createElement): production
calls -2.6% to -7.7% per operation, DOM mutations unchanged, paired timing
within noise. The loops iterate with for-in so descriptor updates allocate no
key arrays (an Object.keys version measured +3% to +12% slower).

Adds a spread/descriptor parity test; the first unification, which re-applied
an unchanged raw-HTML object on descriptor hosts, fails it.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

…from their creators

childSlot named renderPortalState, the implicit signal-handle binding and
useContextInternal directly, so every app that renders a renderable hole
shipped all three graphs. Each value can only come from its creator, so the
creator now installs the regime as a driver: createPortal installs the portal
child, enableSignalBindings the handle subscription, and context creation the
Context read. childSlot keeps only truthiness tests.

shell_ app (oxc), against the previous commit: full gz 122,291 -> 119,422
(-2,869; raw -9,047), base gz 58,047 -> 55,076 (-2,971; raw -9,362).
js-framework work counters: no increase on tsrx, jsx or the plain
createElement fixture.

Adds a Context-as-child React 19 parity test (no prior coverage) and a
bundle-reachability descriptor scenario that drops all three regimes while
the context and hydrate-frame scenarios prove the oracle still sees them.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

# Conflicts:
#	benchmarks/README.md
#	benchmarks/bundle-size/run-minimal.mjs
#	packages/octane/src/runtime.ts

…undles after the merge

The merged childSlot reached the signal-child binder through a local copy of
the driver, which the minifier does not fold, and bindImplicitSignalChild and
enableSignalBindings referenced each other, a cycle the minifier cannot drop.
Test the driver and the engine-loaded flag directly, and enable the signal
document directly from the binder, as the signal-free gate already did.
Signal-free shell base: 56,518 -> 55,076 gzip.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

Co-Authored-By: Claude Opus 5.5 <[email protected]>