frames: C2 — live wire tier

[ryansolid]

Summary

Frames savings pass §3 row C2 — the live wire tier, in the lean form the maintainer ruled, on residue 4's head (55eb4c346, the stack #3867 → #3866 → #3864 → #3863 → #3862 → #3860 → #3849 → next). The live wire moves out of the eager frames client into the sixth lazy chunk, @solidjs/web/frames/wire (frames/src/wire-tier.ts → wire.js, 1,922 min / 934 br):

• the connection (connect): the per-address connection a live() loop's response runs under — the open-frame count its end is judged by (an end with frames open is a death the loop reconnects from; with none a completion), the join of a second loop onto an address already live (one wire per address, the second loop reads the same store), connection.ended / cancel, the SSE reader selection (wire.open(body) for an event stream, ChunkReader otherwise), proxied so a drained body applies nothing further;
• supersession (cancel), reached from bump through tierLoads.wire?.r?.cancel;
• the have-list ledger per mount (have / haveOf, RFC 11 §9.5), fed by the client's #recordHave at its three sites, with encodeHaveList and the FRAME_HAVE_* literals bundled;
• the resume shape (resume → { position: String(version), headers: { "X-Frame-Have": … } }).

The eager client keeps: applyFrames reading options[LIVE_WIRE] as its reader (one line); handle's live arm — prepareTier("wire").then(() => tierLoads.wire.r.connect(…)) — awaiting residency before the body is read, so a live connection without the tier cannot happen; onLive() { prepareTier("wire") }; resume delegating.

Preload-at-call, as the plan's §1 wire row bounds it: live()'s makeIterable fires responseHandler.onLive() at the call, before its first fetch — the import races only the request; a non-live call never loads the tier. The hook rides responseHandler.onLive (the maintainer's adjustment 2 — no new ServerFunctionsClientConfig key; measured ≈ 45 br lighter on the pages).

Running number: frames eager 33,767 / 11,193 → 32,802 / 10,888 br — 888 B above the ≤ 10 KB target (was 1,193 B above). What remains of the incremental path is C1 (≈ −440) and the claims chunk (≈ −270).

Not in this PR: the trace update-site wait (landed apart, #3861).

Measured before written

Re-measured on this head (the base had moved by residue 1–4 — the sync loop, insert-placed fills, the table-owned pending read — so the phase-1 edits were re-derived: .wt-logs/c2b-edit.mjs, c2b-measure.mjs; edited dist copies of the frames client and the sf client through scripts/size's own bundler). min / br; frames eager, then page base, page live:

variant	frames eager	page base	page live	non-SC
base (55eb4c346)	33,767 / 11,193	112,938 / 36,071	124,810 / 39,747	—
del — the ceiling (the wire out whole; unsound alone)	−1,432 / −472	−1,404 / −483	−1,399 / −464	0 / 0
lean — the tier owns the connection, the ledger, the resume; the client keeps the arm, the hook, the delegations	−965 / −310	−979 / −285	−946 / −300	0 / 0
the carrier (lean − del)	≈ +467 / +162
of which the hook (onLive fired at the call)	≈ 20 / 15		+85 br
the ledger's residue (#recordHave ×3 forwarding)	≈ 112 / 27
the field residue	≈ 106 / 14

Net −310 br on frames eager cleared the −200 stop line, so the tier was built. Built: −965 / −305 (+0 min / +5 br over the copy).

Public API changes

1. Removed: Frame.have?(): Record<string, string> | undefined from the public Frame interface (frames/src/frame-client.ts). The ledger is the tier's haveOf(frame) (@internal). The maintainer's adjustment 3 — an approved-class removal. frames-live-resume.spec re-pointed.
2. Moved: encodeHaveList from frames/src/frame-transport.ts into the tier (@internal). decodeHaveList and the FRAME_HAVE_HEADER / FRAME_HAVE_BUDGET constants stay in frame-transport (they were already off the client entry — residue 1, frames: residue 1 — D list, FRAME_HAVE_* off the client entry, lean re-ask #3863; nothing to re-export).
3. Added: the @solidjs/web/frames/wire export path (the chunk; types/frames/wire-tier.d.ts; every export @internal).
4. Added: ServerFunctionsClientConfig.responseHandler.onLive?(): void — optional; fired by live() at the call, before its first fetch, per call. ServerFunctionsClientConfig itself gains no key.
5. Added: ServerComponentHandler.onLive(): void (the frames handler's implementation of 4).

Behaviour, no surface: while the tier is not resident, responseHandler.resume answers undefined (the ledger lives in the tier) — see the degraded case under Pins.

Pins

consistency/tier-wire-preload.spec (new, 3): the hook starts the import synchronously at getRoom(), before any fetch; the response arm awaits residency (chunks sent, nothing applied, the fallback shown, no status) and the release lands the content; X-Frame-Tiers: wire on a live response joins the same load (the loader called once); a non-live call never loads the tier and its refetch bumps nothing of the tier's; reconnect through the tier carries Last-Event-ID and the have-list ({ "": root, "lh:1": hole }), morphs in place, completes to closed; the degraded first connect, pinned as accepted: content a plain call applied with digests while the tier was gated → the live call's first connect carries neither X-Frame-Have nor Last-Event-ID → full snapshot; the release lands the live root; the reconnect after a death carries both from the live root's ledger.

The brief described the degraded case as "sends the ordinal with no have-list". With resume fully delegated (the lean form) the ordinal goes too while the tier is absent. The server never read it for a frame render — liveSource compares Last-Event-ID to a value digest a frame answer does not have; only decodeHaveList(X-Frame-Have) is read in the frame sink — so the outcome is the same full snapshot. Pinned as built; flagged here in case the ordinal was wanted for its own sake.

Every live suite green through the tier: frames-live-resume re-pointed to the tier's haveOf / encodeHaveList; frames-live and frames-live-showing warm it (beforeAll(() => prepareTier("wire")) — resident cells); frames-get, frames-plain-bound, the server frame-live-* / server-functions-live-* suites unchanged.

The coverage closure (server-functions-live-loop.spec, new, 19, client suite): live()'s loop driven from the consumer's side over hand-framed bodies — the codec's chunk stream, and the same payloads as server-sent events with id: lines — ending as each test says (cleanly / dropping / hanging until the request's signal): the non-server-function throw; a handler without intercept (and onLive fired before the fetch); a first-connect failure as the call's rejection (no status, no retry); a death reconnecting on the online wake, the reader's cursor riding back as Last-Event-ID; the fail-fast 4xx closing with the error; 408 / 425 / 429 retrying; Retry-After naming the wait; return() mid-connect (the arriving stream closed, the wire severed) / mid-backoff (the sleep cut) / mid-seed (a late local answer not yielded); invoke's signal mid-stream (a rejection, ended for good) / mid-backoff (the sleep woken, the connect refused); a host without a global EventTarget; live(GET(fn)) at the live address over GET; the handler's local answers (a hit at the call yielded first and adopted; a deferred hit awaited, settling to a value / to nothing; a miss at the call but a hit at connect — one value, no wire).

live's per-function branch coverage, counted as the audit did (§2.4: v8 coverage over the client suite + the hydrate suite, merged; server-functions/src/client.ts, the function live):

	branch sites fully covered	branch paths
base 55eb4c346 (its own tests)	26 / 59	91 / 129
this head	57 / 60	129 / 132

The three left: return() landing between a value's arrival and its read (L1593), and the resume / seed arms' stopped-while-pending elses (L1527, L1579's implicit else). The brief's "22/61 → ≥ 45/61": the audit's 22/61 was counted on an older head whose live had 61 sites; by sites on this head the closure is 26 → 57 of 60.

Suites: web client 130 / 1,218; hydrate + consistency 89 / 452; server 159 / 1,511; solid 42 / 833; types + test-types clean. Harness: CONSISTENCY_CASES=500 × seeds {3289, 91501}: SC arm 0 findings; generic arm (CONSISTENCY_IGNORE=C1,C9,C19,E) 0 findings.

Size

scripts/size on this head (fresh turbo --force build, native compiler built fresh), vs the base 55eb4c346 and vs next @ 9d89df731 (unchanged since int3's fresh measurement). min / br:

scenario	this head	vs base	vs next	cap
frames: eager client consumer	32,802 / 10,888	−965 / −305	−10,612 / −2,899	11.21 → 10.90 KB
page: base server components	112,003 / 35,722	−935 / −349	−33,754 / −9,160	36.09 → 35.74 KB
page: live server components	123,903 / 39,412	−907 / −335	−33,817 / −9,183	39.76 → 39.43 KB
the other 14 scenarios		0 / 0	0 / 0 (hydrating ×3 the base's A0 exceptions; server renderToString +4 / −9 within ±10)	unchanged

• Chunk, reported not counted: wire.js 1,922 min / 934 br (the plan's est. ≈ 1,300 / 470 priced the dispatch; the tier carries the whole connection). The other five lazy chunks unchanged.
• Live page, honestly: this page LOADS wire.js at its first live() call, so its total transfer is 39,412 + 934 = 40,345 br against the base's 39,747 (+598) — the chunk's own weight plus the carrier, the price of the eager −335 and of the frames-eager −305 on pages that never go live.
• Caps ratcheted to measured + 10 B at the 0.01 KB step (scenarios.js ledger lines, floor-caps.json with recorded minified); no cap raised (check-floor-caps.mjs against the base: clean). The three pre-existing local FAILs (signals: core floor +11 B, + isPending/latest +23 B, CSR with Show/For +45 B over cap) are byte-identical to every earlier head and to next — the known local-vs-CI brotli difference, not this PR's.

Not done

• live's last three branch sites (above) — a return() race and two stopped-while-pending elses; not reachable deterministically from the consumer's side.
• The ordinal on the degraded first connect (flagged under Pins) — the lean form drops it with the have-list; a two-line resume fallback in the client would restore it at ≈ +40 br if wanted.
• C1 (holes tier) and the claims chunk — the remaining items of the incremental path to ≤ 10 KB.

[changeset-bot]

🦋 Changeset detected

Latest commit: d9395e3

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

This PR includes changesets to release 12 packages

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

Folded into #3860 (wip/frames-tiers-integration) per the maintainer — "merging is a bottleneck." The stack #3860 → #3861 → #3862 → #3863 → #3864 → #3866 → #3867 → #3868 was linear (every head an ancestor of the next), so the integration branch was fast-forwarded to this PR's head d9395e398; this PR's own commit is the top of that chain unchanged. The branch is kept.

On the integrated head, built fresh: frames eager 32,802 min / 10,888 br (−2,899 br vs next @ 9d89df731), page base 112,003 / 35,722 (−9,160), page live 123,903 / 39,412 (−9,183); wire.js 1,922 / 934 lazy. Caps as this PR ratcheted them (10.90 / 35.74 / 39.43 KB); check-floor-caps.mjs against the Phase A base: no cap raised. Suites, test-types, the dist specs and the 500-case harness (both arms, seeds 3289 / 91501) green. The degraded first connect (no have-list, no ordinal) and the ≈ +40 br ordinal fallback are #3860's decision 17.

— Claude via Cursor