Commit cbc3d20

"@solidjs/signals": patch

A pass's verdict-reader posture is toward the hold it was earned against, not toward every hold at once: the nodes a probe asked about are remembered per node and resolved to the holds that hold them at use (a lane's node to the hold its lane is under, a derived store's leaf through its family), so a later separate hold sees the pass as whatever its reads make it, and the posture lapses when its hold lands. At the head, a held node seated as the flush's work whose probes name nothing any live hold holds is that hold's work, not its verdict reader (#4417, #4544, #4574).

**Hold identity (as built, step 4, 2026-11-01).** `probesHeld(c, toward)` compares _holds_, not levels: `holdOf(p)` resolves a remembered node to the hold that holds it now — `holdOfTx(p._x._transaction)` for a held or override node (`resolveTx`, then up through `_lane` parents to the real hold), through `_fam.node` first for a derived store's leaf (D-30), or, for a node staged this flush or in flight and not yet held, the flush's hold to be (`flushTransaction`, or a `FLUSH` sentinel when the flush has not made one yet) — and `toward` is resolved the same way (a node, a transaction, `null` for the flush's own proposal, `undefined` for "any live real hold"). Ruling 1 (per hold, conditional on isolation): toward a later, separate hold the posture earned against T does not apply; a D-4 merge makes the two one hold and the posture carries. The head clear (Ruling 2, today's order): a `CONFIG_HELD` node seated as `tx`'s work at the head whose probe memory names nothing any live hold holds drops `CONFIG_VERDICT` _before_ `joinPassTx` re-keys this flush's stagings to `tx` — a pass that probes T and merely stages for T is T's verdict reader; only a fetch T waits on (#3884) or a node born into T (D-27) stays T's work against its own probe.

**Status after step 4 (2026-11-01, branch `fix/seat-per-hold`).** Built: steps 2–4 (shadow, read side, head with the `8ccab0f71` provisional re-derive and the hold identity above). Gates met: #4417 `[1,2,2]` ×4 (signals and web) pass; the 49 isolated rows keep today's screen; full signals, web (client/server/hydrate) and solid suites green; the async-gen regression gate (seed 3289, 1500 cases) reports no regression and no improvement. **The targeted replay does not pass**: of the 1217 saved ids (seeds 7 and 3289, the `_b*`/`_cb*` lists) 33 differ from `39f951b84` (`/tmp/gen/seat-tracediff.txt`), and bisecting the branch's three layers (read side `5417006de` / `+ 8ccab0f71` salvage / `+ head`) puts them in two groups, neither forced by §2:

- **Read side (committed `5417006de`) — 15 ids, two of them liveness failures** (7 #6390, 3289 #30951: `[ROOT-L]` forever; the others 7 #21333 #25272 #29429 #43308 #44018, 3289 #840 #22719 #24851 #25679 #29047 #35194 #43096 #47870). Reduced (`/tmp/gen/seat-liveness-repro.test.ts`, signals level, passes on base): memos created _outside any root_ (the harness's un-owned form) `m0 = async(s1)`, `m1 = async(s1)`, `m2 = async(m1)` where `m2` rejects, and under a `Loading` a render effect reading `[isPending(m2), m0()]`. When `m2` rejects the effect re-runs, its read of `m0` finds it stale and recomputes it (both base and head re-ask `m0`); base serves the committed `3` through `observeFlight` (node-level `CONFIG_VERDICT`), head's `probesHeld(c, m0)` says the effect has no posture toward `m0`'s flight (its memory names `m2`, errored, held by nothing; an ownerless flight has no transaction, so `holdOf(m0)` is the `FLUSH` sentinel) and lets it throw `NotReady` — rule-faithful so far (Ruling 1) — but when `m0` lands the effect is never re-run: `propagateStatus` (async.ts:1099–1108) still consults the node-level bit and treats it as a verdict reader to mark, not re-derive. The §3 split (node bit for `propagateStatus`, per-hold seat for reads) is inconsistent for a plain pending read made by a node that carries the bit. Not fixed.

- **Salvage `8ccab0f71` — 18 ids, F-2 empty slots** (7 #1596 #1866 #4423 #13985 #21348 #24789 #30013 #41189 #45702, 3289 #1831 #8668 #12357 #12695 #17777 #24610 #30977 #31700 #36682; a few also stop breaking an E3 law base broke): under `<Show when={isPending(m0)}>` a leaf's `m0()` text hole (or attribute) stays empty where base shows the value, e.g. `[s0=1 m0=async(s0 !%4=3)] <Loading><Show when=isPending(m0)><span>m0()</span></Show></Loading> :: S4 t S2 s0=3 m S3 s0=0 S3` → `<span></span>` from step 5 on (base `vm0:2`). The in-place re-derive adopts the readers that derived from the staging; here the adopted `insert` effect is not re-run at the seam. Not fixed; the salvage's own two cases (`adoption-3289-464-91501-411.spec.ts`) still pass. The head layer (this PR's per-hold seat and head clear) moves none of the 33 on its own.

Three findings from the gates themselves, each pinned `it.fails` at the ruled screen rather than fixed:

- **S-46 web plain row** (`display-ahead-plain-write-content-s46.spec.tsx`): §4's mechanism ("`Show`'s render effect that attaches the branch is created by the same pass and reads the born-held node") does not match the real structure. The attaching effect is `insert`'s, created at mount, reading `Show`'s value memo P — T's verdict reader, published display-ahead; the hole's `insert` effect inside the branch _is_ born held (correct), but nothing seats P's output as T's, so the branch still attaches with its hole empty. The only mechanism found (dropped `04c218537`: a born-held uncaught render effect marks its creator pass joined) looped on generator 91501 #32043. Needs a mechanism ruling.

- **#4544 Shape 1 probe rows** (`verdict-memo-held-branch-4544.test.ts`, class (c) rule conflict): on the `plan` write's flush the join makes `plan` T-held; `chrome`'s D-22 provisional probe re-derives at the seam as T's verdict reader; `params` (T-held, inheriting `[plan]` by #4050) is seated T's verdict reader at the seam re-pass — Ruling 2 as pinned by `latest-first-one-pass-4402` covers inherited postures — and is served the committed `session`; the born-held branch asks `signedOutOnly` and T never lands (`0 signed out`). §4's path (row 3 "work beats verdict" for `chrome`) was narrowed away by §5 to keep Shape 2.

- **Timing edge, the lane's side** (same file, last row): `params` is U's verdict-lane node; the pass T's write dirties is seated as U's lane work at the head (`setPassLane`: `CONFIG_OVERRIDE` without `CONFIG_GUESS`) and stages into the one lane slot, so U's readers see `signedIn: true` before T lands. D-37 as amended wants T's staging beside the lane's value — the two-slot seat named under "Relationship to #3905", not built here. Base behaves the same.

Size (step 5, measured against `39f951b84`): `+ isPending/latest` +670 B br / +2085 B min (verdict.js +884, lanes.js +683, core.js +365, scheduler.js +141 minified); live page +717 B br / +2160 B min; hydrating + every store family +202 B br / +860 B min; core floor +22 B br / +70 B min. Over the +150 B br line — stopped before the PR; the breakdown is in the step report.