@@ -887,3 +887,42 @@ discovery constant — solid server signals/boundaries were 24% of shell
887887CPU (read/runWithBoundaryErrorContext/recordSlot/NotReadyError throws),
888888GC ~ 25%. That is §15 territory: boundary-setup + per-read costs, same
889889family as the news per-element gap.
890+
891+ ### 15. The discovery constant, decomposed (2026-08-23, morning)
892+
893+ Attempted spot-fixes, with the measurements that killed or kept them:
894+ - Error.stackTraceLimit=0 A/B: DEAD EVEN (101 vs 101 µs) — NotReadyError
895+ stack capture is NOT a cost; do not "fix" it.
896+ - FormData-plugin frames in heap profiles are one-time undici lazy-loading
897+ at startup, not per-render — discount loader frames before reading.
898+ - Slot flight-memory (#3003 ) rebuilt: WeakMap<ctx, Map> -> symbol-keyed
899+ plain object ON ctx, records mutated in place (single {s,v,d} shape).
900+ recordSlot was 12.7% of recurring allocation; it leaves the top-12 —
901+ but µs/shell does NOT move. Young-gen garbage at this scale is free.
902+ - ssrLoadingBoundary: parent contexts read without owner switches;
903+ reveal-group severing setContext (a context-map clone) now only paid
904+ when a group is in scope. Under noise floor. (All landed: d8db8eb3.)
905+
906+ THE DECOMPOSITION (the §15 result): a floor variant of the same page with
907+ every promise pre-stamped {s:1,v} — full content, zero async machinery —
908+ renders in 57 µs. Octane's ASYNC shell is 60. Our staggered shell is ~ 107.
909+ The sync render path is ALREADY at octane speed while emitting MORE HTML
910+ (9.7KB content vs 4.2KB fallback shell). The entire remaining shell gap is
911+ the per-pending-boundary constant: ~ 5µs × 10 boundaries = discovery run +
912+ NotReadyError throw/catch + deferred + serialized stub (~ 10µs total for
913+ the batch session, the largest single slice) + registerFragment + boundary
914+ owner/context ceremony. Cost is smeared across ~ 30 frames of 1-4% — a
915+ structural constant, not a hotspot.
916+
917+ WHAT WOULD ACTUALLY CLOSE IT (design questions, not spot fixes):
918+ - Stub-less shell protocol: today a record's PRESENCE in _ $HY.r at hydrate
919+ time is how the client learns "server owns this async value" —
920+ sharedConfig.has() returns false for an unarrived record and the client
921+ computes fresh. A client-side deferred FACTORY (boundary registration
922+ creates pending records on demand; fulfillments create-or-resolve) would
923+ delete the whole shell stub session, but it rewrites the has()/load()
924+ contract and every consumer that branches on record presence.
925+ - Lighter boundary discovery: the ceremony (buffered ctx via
926+ Object.create, ~ 10 closures, error-context save/restore per phase) is
927+ per-boundary; a compiled or pooled boundary form could amortize it.
928+ Both belong to a protocol conversation with Ryan, not a solo increment.
0 commit comments