[ryansolid]

Step S0 of documentation/plans/reader-seat-rebuild.md (brief: reader-seat-brief.md §4). Draft — not for merge. The swap (S1) and the cleanup (S2) are separate PRs, gated on the rulings listed under Two readings below.

What is shadowed

packages/signals/src/core/seat.ts is the new layer: the seat (openSeat — world S < Fl < H that only raises, lane, display-ahead, posture from the node alone; bornHeld — row B2) and the slot table (slot(c, el) — rows U1–U4, F1, A1, L1–L8, D1–D8, V1–V6, Vl1–Vl5, Vp1–Vp2, with the primed rows the differential added). Everything is behind __TEST__ ("false" in every rollup define and size script).

Under __TEST__:

• every read runs once for real and slot() predicts beside it at the deciders' decision point (fast block, slow path after the pull, verdictValue's window, laneRead); each return site of the old deciders records the slot it served (shadowActual). Compared: slot identity (C / P / Lv / Dv / T / NR / E), and the reader's world state after the read for non-leaf postures (INV-seat-3).
• every recompute opens a seat beside the head seating (setPassLane / laneRoot / mount) and compares lane, ahead and inherited world (row B1), and in the finally the born-held decision and the lane the pass ends in (row B2). A leaf never raises its world (INV-seat-1); posture is recomputed from the node at open and close (INV-seat-2).
• raises the old code applies outside read (joinPass/joinPassAt, stagedRead, the boundary fallback swap) are mirrored into the seat.

Mismatches are recorded, not thrown (__SEAT_SHADOW_STRICT__ throws): SEAT_SHADOW_REPORT=<file> makes the signals/solid/web vitest configs append one JSON line per test file; packages/signals/scripts/seat-shadow-summary.mjs aggregates. SIGNALS_TEST_DEFINE=true pnpm --filter @solidjs/signals build emits a dev dist with __TEST__ on so solid, web and the generator (which consume dist/dev.js) run the shadow; the default build is unchanged.

Docs: design doc §3 (harness as built), §4.2 (results), §6 (table rows found wrong against a rule — fixed in seat.ts, the brief's table untouched), §7 (two readings); INTERNALS-ASYNC-STATE.md §5 gains the INV-seat family; invariants.ts points at it.

Suite results

run	reads compared	passes seated	mismatch classes	instances
signals suite (376 test files; re-run with the table as committed)	965,386	238,343	40	151
visibility oracle (3 files: posture / oracle / store)	11,379	8,322	0	0
solid suite (49 test files)	2,389	1,546	0	0
web suite (27 .spec.ts files; the 153 .spec.tsx files fail to load in this worktree on next too — no @solidjs/compiler binary)	254,902	155,740	25	74
generator, 2,000 cases × seeds 3289 / 91501 / 7, every pair (E1 E2 E2e E3 E6 E7 E8 E9, owned + unowned via E7)	6,969,627	5,153,328	143	4,127
generator, 50,000 cases × seeds 3289 / 91501 / 7, every pair — table before the two 50k fixes below	67,328,944	43,755,988	220	8,821
generator, 50,000 cases × seeds 3289 / 91501 / 7, every pair — re-run with the table as committed	67,328,944	43,755,988	209	8,706

Every suite is green with the shadow on. The 50k run is deterministic (same reads/passes on repeat; the re-run with the committed table loses exactly the 11 classes / 115 instances the two fixes cover and gains none) and added two table rows (L4/L2/V1' staged error → E, D-41; V3 under published inputs in flight → the flight's V4', F-9 / #3305) and no new two-readings item; the §7 buckets scale roughly linearly from 2k to 50k (F-ahead-posture 1,461 → 2,922; F-vlane 1,054 → 2,355; F-vworld 1,101 → 2,308; F-transitive 301 → 684; B2 display-ahead residue 104 → 257; F-mount 97 → 180).

Generator findings on next, not shadow findings (reproduced on the plain build without the shadow; not minimized or filed here):

• all three seeds abort before 50,000 cases with Potential Infinite Loop Detected. Kept alive by scheduled work; last staged node: … — seed 7 inside the first 2,000 cases (computed), 3289 at ~427 s (value), 91501 at ~638 s (signal). Verified against a dist built without SIGNALS_TEST_DEFINE (no shadow in the bundle): each aborts at the same case index on both builds — seed 7 at case 1158 (in the E8 remount run), 3289 at case 3963 (the base run), 91501 at case 5807 (the E8 remount run). The campaign does not print the index; a throwaway replay of its survey loop (same fc.sample stream, same per-case run order) did — not committed. Repro: ASYNC_GEN=1 ASYNC_GEN_SEED=<seed> ASYNC_GEN_CASES=50000 ASYNC_GEN_MODE=survey ASYNC_GEN_PAIRS=E1,E2,E2e,E3,E6,E7,E8,E9 pnpm exec vitest run test/async-gen/campaign.spec.ts in packages/web.
• seeds 3289 and 91501 each report one equivalence/E7 signature (owned vs unowned screens) in survey mode, on both builds.

Mismatch triage

No class was found where the NEW slot is right and today is wrong by a rule, so no issue was filed. Every slot disagreement traced either to a table gap (the table was fixed in seat.ts; design doc §6) or to one of the two-readings items below (design doc §7), which need a ruling before the swap. Counts are signals / web / generator-2k.

row(s)	shape	which side	rule	status
U4	untracked read of a memo with _shownError (write/retry over an error)	today (E)	D-41, #4110	table fixed
Vl1'	latest() by a render effect outside the hold's flush, verdict lane blocked	today (C, stale reader)	#3460 / A15	table fixed
Vl1, Vl4, Vp*	staged error inside latest throws; isPending never throws, error last	today	D-41	table fixed
L6, D4'''	lane work reading a lane derivation in flight	today (NR)	D-37	table fixed
L7'	leaf held by the flush's tx reading a lane derivation's staging	today (P)	D-2	table fixed
V6	a verdict pass's derivation (today a verdict lane's CONFIG_OVERRIDE value)	— (new row)	D-25, #4050	row added; residue under F-vlane
B2 leaf clause	render effect under a verdict-lane mount reading a staging is born held	today	D-27, #4166	table fixed (mark SR)
B2 exclusion	nested-derivation-left-pending exclusion vs a verdict reader's direct mount	today	#4203 #4340 #3851	table fixed (creatorDirect)
D1 / L1	staged + uninitialized + in flight → NR	today	F-9	table fixed
D2	held, nothing staged → C with the join	today	D-2	table fixed
L4 / L2 / V1'	leaf or born-into frame reader of a held node whose staging is an error	today (E)	D-41	table fixed (50k)
V3	verdict reader's plain read of a node staged + in flight with INPUTS_PUBLISHED	today (NR via V4')	F-9, #3305	table fixed (50k)
F-vlane — V6 / V4' / V1 / V2 / L2 / Vl1' under oldVlane	a node whose staging sits in a verdict lane is seated as the verdict lane's on its next pass (incl. pending re-pass after #3884 clears CONFIG_VERDICT, and the #3540 boundary swap); the brief gives verdicts no node set	two readings	D-25, #3884, #3540, #4337	15 / 11 / 1,054 instances; ruling needed: how posture persists without a node set
F-leaf — B1 inherited world, B2 born held (old true, new false, world 0)	a render effect that read a born node is JOINED today and its children are born held; INV-seat-1 keeps a leaf at S	two readings	A29 vs D-1 / INV-seat-1	41 / 0 / 0
F-mount — B1 ahead (old true, new false; creator running false)	today's display-ahead mount is module state (laneRoot), covering runWithOwner creations (boundary fallback memo); the brief says creator = context	two readings	A29 "children"	22 / 3 / 97
F-ahead-posture — B1 ahead (old false, new true)	brief: ahead = creator verdict posture or lane; today only a lane/passLane	two readings	D-27	1 / 29 / 1,277 (+184 the #4341 narrowing, V in world ≥ Fl)
F-vworld — equal slot, world after the read old 1|2 new 0 (Vl/Vp/V, D5)	verdict reads set STAGED_READ/JOINED on the reader today; the table gives marks only	two readings	D-1 vs D-23/A10	40 / 25 / 1,101
F-transitive — B2 born held (old false, new true; world 2, ahead false)	world inherited transitively through running creators vs today's one-level creatorPass(oldcontext) JOINED	two readings	A29	14 / 2 / 301
residue	L7 Dv vs Lv + lane seat at pass end (spec-async-semantics ×2); Vl1' C vs P with the reader in world H (generator 50k ×10) and V1' P vs C for a render-effect verdict reader born into the hold (stagedScreen arm, 50k ×2); B2 under a display-ahead mount (generator 50k ×257, 36 classes)	untraced	—	listed in the summaries, not classified
residue, 50k (literal reading)	the 50k classes (three seeds; unchanged by the two 50k fixes) that match no §6 finding or §7 item when the class signatures §7 names are read literally — 623 instances, 43 classes: B1 ahead (old true, new false; creator running true; creator memo:value / children / computed, effect) ×406, 4 classes (the previous triage filed these under F-ahead-posture; §7's text names only old false → new true); D5 E vs NR for an errored node in flight (u f · f ip · u f ip · one f d vlane; D and V postures; 4 classes under oldVlane) ×101, 17 classes; L8 E vs NR, the same shape at a leaf ×34, 6 classes; V1 C vs P (s h, oldNoLane, t1 w0) ×57, 2 classes; V4 C vs NR (f ip; ×8 in 3 classes under oldVlane, ×4 in 2 classes oldNoLane) ×12, 5 classes; V1'' P vs E (s h u ip — the D-41 staged-error shape on a row the L4/L2/V1' fix did not touch; V1'' still serves P for an uninitialized held node) ×2, 2 classes; V1' P vs C (s h, w2) ×2, 2 classes and Vl1 P vs C (s h, w0, oldNoLane) ×1, 1 class; world after the read old 2 → new 1 with equal slots (Vl1 P ×4, Vp1 C ×2, Vl5 Dv ×1) ×7, 3 classes; B2 born held (old true, new false; world 1, ahead false, mount false, rows L8) ×1, 1 class	untraced	—	listed in design doc §4.2, not classified

Design doc §7 has both readings and the screen each produces for every F-item.

Size

pnpm build after the change; scripts/size head vs the next base dists (diff -rq of packages/{signals,solid,web}/dist against the pre-change snapshot shows solid and web byte-identical; signals differs only in the prettified dev/observe/prod sources where if (x) return y; became if (x) { return y; } around a dropped __TEST__ statement, plus the new types/core/seat.d.ts). Minified and brotli bytes are identical on every scenario that bundles:

signals: core floor (createSignal/Memo/Effect/Root/flush)  brotli   7413 minified  19971 = base
signals: + createStore                                     brotli  14924 minified  43916 = base
signals: + isPending/latest                                brotli  10379 minified  28678 = base
signals: + until/refresh (truth readers, no optimism)      brotli   8037 minified  21776 = base
app: render + one signal (the simple-app floor)            brotli   9769 minified  26803 = base
app: hydrating (no stores) with Show/For/Loading/Errored   brotli  18606 minified  53357 = base
app: hydrating + every store primitive family              brotli  31758 minified  96961 = base
app: CSR with Show/For/Loading/Errored/lazy                brotli  13380 minified  36932 = base
app: CSR, observe tier                                     brotli  15105 minified  41880 = base
app: CSR, observe tier + attribution engine enabled        brotli  29713 minified  88219 = base
frames: eager client consumer                              brotli  11754 minified  34886 = base
page: base server components                               brotli  33533 minified 101358 = base
page: live server components                               brotli  37954 minified 115118 = base
server: floor (getRequestEvent + isServer)                 brotli   1275 minified   3255 = base
server: renderToString (the server-render floor)           brotli  22786 minified  77473 = base

The seven scenarios that compile JSX (five compiled-template, two router) fail to bundle on both head and base in this worktree (@solidjs/compiler-wasm32-wasi binary not installed) — an environment gap, not a size result; CI measures them.

Public API changes

None. No export, prop, option, parameter, default, documented behaviour or diagnostic code changes. SIGNALS_TEST_DEFINE is a build-time environment switch for the rollup config, SEAT_SHADOW_REPORT / SEAT_SHADOW_TRACE / SEAT_SHADOW_STRICT are test-harness environment variables; none reach a published artifact.

---

Co-authored-by: Claude via Cursor

[changeset-bot]

🦋 Changeset detected

Latest commit: 697217f

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 15 packages

Name	Type
@solidjs/signals	Patch
test-integration	Patch
@solidjs/web	Patch
@solidjs/babel-plugin	Patch
@solidjs/compiler	Patch
@solidjs/diagnostics	Patch
@solidjs/element	Patch
@solidjs/h	Patch
@solidjs/html	Patch
solid-js	Patch
@solidjs/universal	Patch
board-example	Patch
canvas-example	Patch
server-dbmon-example	Patch
todos-server-example	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[ryansolid]

Superseded by #4438, which lands this S0 work together with the S1 M1 re-baseline (4c413e1ea: the shadow against next @ 5e68925, 0 mismatch classes on every suite and on 2k × 3 seeds, shadowKnown for #4426 #4427 #4428 #4429, dist byte-identical).

The swap that was to follow (#4430) is stopped; the record is the design doc's §8 on the seat/spec branch. Closing this draft; the branch seat/rebuild-beside stays.

— Claude via Cursor