frames: residue 2 — the preview pull form, measured and stopped (carrier +483 br over a −546 ceiling; net −63); C15 (e) pins #3844's gap

[ryansolid]

Draft — frames size pass, residue step 2: the preview pull form, MEASURED AND STOPPED. Off size/frames-residue-1 (#3863) @ fce81d2c0; retargets as the stack merges (→ #3862 → #3860 → #3849 → next). Ground truth: documentation/plans/frames-residue-pass.md §3.1 (the design) and §2 row 1c / §5 row 3; #3844's body ("the gap"); frames-rulings.md 1.6 / 2.3 / "Fix shape — seam 1"; principles §9.2.2 (A29). The step's standing rule was measure before writing — build the carrier on an edited dist first; if it exceeds +300 br over the deletion ceiling, stop and report the design. It does: +483 br. Nothing of the form reached source. This PR carries the measurement (docs) and one pin.

Summary

Frames eager stays 35,250 min / 11,654 br — 1,654 B above ≤ 10.0 KB. The pull form was budgeted at −300 br net (a −542 deletion ceiling, a +242 carrier sketch). Measured on this head as the step would have written it, the honest carrier is +1,414 min / +483 br over a −546 ceiling → net −231 / −63 (pages −29 / −37; non-SC 0 / 0). Two findings:

1. The sketch is unsound where its budget was built. pullcarrier (the measured +242) deleted the content token with the buffer and read host.staged(address) off the bare address. But dynamic delivers a kept resolution only when its .address differs from the last one delivered (web/src/index.ts resolveBinding / sameInstance): a refetch of the address a mount SHOWS resolves to the same object, the delivery is a same-value write, and neither the follow effect's compute half nor any fill memo runs in the transaction — the content lands at body end beside siblings the transaction holds. That is frames: A1b + A4 — stage/preview deletion, declared slot records, S-ref #3844's gap (1) restated; the sketch could not have kept C15 (a, b) or any multi-flight arm of frames-optimistic-hold green. The token binding stays in any form: 227 min / 88 br, the first thing the +242 did not count.
2. The honest carrier is twice the sketch, and for a structural reason: the handler's chunk buffer (≈ 200 min) is replayed through host.apply and inherits every semantic the host already has — version policy, the landing, L1, per-id routing, so nested regions and a flight's several roots stage for free. A host-side staged set has to carry each of those itself (promote through the store's own write + landing tail, the write / settleWait guards, flight-path per-root staging 176 / 54, nested regions promoted with the root 152 / 53), and the pull read needs a lane gate — the fill keyed by the mount's own address accessor, so a write-through recompute between a refetch's header and its commit (a live op on an adopted boundary, a reveal cascade, a settleWait) does not read the staged set outside the pass — 106 / 29, which the push has by construction. So the pull read (≈ 750 min) is not smaller than the push it replaces (≈ 700). The unsound floor without flight / regions / gate is still +980 / +339.

Measured before written

edited dist (this head's frames/dist/client.js, through scripts/size/bundle.mjs)	frames eager min / br	Δ vs head	carrier over the ceiling
head fce81d2c0	35,250 / 11,654
ceiling — stage whole out: the buffer, the token, stagedContent / contentAddress, FrameImpl#preview / host.preview, the token arms of handle / applyFlightResponse, followAddress → a rebind effect	33,605 / 11,108	−1,645 / −546	—
the carrier as designed (below)	35,212 / 11,651	−38 / −3	+1,607 / +543
… promote through write + the shared landing tail (the lean form; sound)	35,019 / 11,591	−231 / −63	+1,414 / +483
lean − the token binding (unsound)	34,792 / 11,503	−458 / −151	token = 227 / 88
lean − flight-path staging (unsound)	34,843 / 11,537	−407 / −117	flight = 176 / 54
lean − regions (unsound)	34,867 / 11,538	−383 / −116	regions = 152 / 53
lean − the gate (unsound)	34,913 / 11,562	−337 / −92	gate = 106 / 29
lean − all three (the unsound floor)	34,585 / 11,447	−665 / −207	+980 / +339

Pages (lean form): base 114,422 → 114,184 / 36,488 → 36,459 (−238 / −29); live 126,294 → 126,061 / 40,125 → 40,088 (−233 / −37). The four non-SC scenarios 0 / 0 on every row. Tooling: .wt-logs/res2-edit.mjs (the cuts, exact-string replacements asserted to match once; node --check on the result), res2-measure.mjs (the harness's own bundler with the copy in place of the dist). Built: none.

The carrier as designed (what "+483" buys; the text measured is the text the source change would have produced). Handler: showing decides at the header as today; a staged response's chunks reach the host flagged (host.apply(chunk, true); the flight path flags by staged-root prefix, as regionOf did); the token binding is minted into latest; the call settles at body end. Host: apply(chunk, stage) writes a flagged chunk whose version is newer than the store's into store.staged = { version, records } under the same policy-A write (one response's; once the store is at the version the rest writes through, as a committed entry did); settleArgs through the store's version-keyed waits, settleWait re-applying only a record that is the store's current one; staged(id); promote(id, version) = the store's own write + the landing tail apply already runs (shown / open / the pending landing settled or rejected / frames applied as one write / L1 for the response's undelivered refs), the waits carried across the bump. Frame: #stagedRecord(occurrence) — the staged set's record, else the parent's when this store does not shadow the key (the inherited rule preview had); ctx.staged() resolves it through #resolveArgs unless it adds or renames a region (R.changed). Fill: liveSlotProps reads staged() && ctx.staged() || args(), staged = host.staged(address)?.version === +version parsed from the mount's binding() — the signal dynamic writes in the transaction's pass, so the read is lane-correct. Follow: the effect half promotes the token's version, rebinds, then promotes each region store below (regions-tier.ts, replacing frames()).

Public API changes

None — no source under packages/*/src changed. For the record, the form would have removed FrameHost.preview / Frame.preview / stagedContent / contentAddress (approved 2026-10-06) and added FrameHost.staged(id), FrameHost.promote(id, version?), a second parameter on FrameHost.apply(chunk, stage?), and SlotContext.staged() (all @experimental); a custom FrameHost not implementing them would lose atomic staging (today the handler's buffer gives it to any host). Named here so the trade is visible if the form is ever taken at −63.

Pins

• Added, by name — c15-staging-atomic (e), the gap frames: A1b + A4 — stage/preview deletion, declared slot records, S-ref #3844 named: a same-address refetch enters the transaction — with the body complete and the sibling held, the fill HAS derived the new arg (seen is ["one", "two"]: the staged read happened in the transaction's pass) while the DOM still shows a0|v1|one, no frame:applied fired and the frame's version is 1; the release lands the whole in one frame; the fill is not re-read at the commit. Any carrier of the staging — the push today or a pull — must keep both halves. Green on the push form (reproduced before any change, per the rule).
• Unchanged and green on this head: C15 (a–c, control), frames-optimistic-hold ×6, frames-morph-in-transition ×3, hydrate + consistency 87 files / 446 passed / 16 expected fail / 2 skipped.
• The pull form's own soundness against the pins was reasoned (the design above), not run — nothing was built to run it against.

Size

No change: frames eager 35,250 / 11,654 br (cap 11.67 KB), page base 114,422 / 36,488 (36.50), live 126,294 / 40,125 (40.14); vs next @ 9d89df731 as #3863's. No cap moved. The three pre-existing over-cap-by-noise warnings (signals floor, isPending, CSR) are +0 as on the base.

Docs

frames-residue-pass.md: status line; §3.1 "As measured" (the sketch's unsound point, the honest carrier as designed, the attribution table, why the sketch was short by half, the conclusion); §2 row 1c and §4 row 3 corrected to the measured net; §5 item 3 marked stopped; §7b the step's table and what it changes in the path — the residue proper stalls ≈ 215 B higher than §4 says; crossing 10.0 still needs C1 and the claims chunk, with ≈ 120 B of margin instead of ≈ 330. frames-savings-pass.md §3: row R2. frames-consistency-contract.md: C15's pin list gains arm (e).

Not done

• The pull form itself — stopped per the rule; recommended against at −63 (§3.1 "As measured", conclusion). The push form stays as frames: A1b + A4 — stage/preview deletion, declared slot records, S-ref #3844 left it. The alternative measured in passing — keep the chunk buffer, pull the fills' read from it (stagedContent.record(token, id, occurrence) + the host's settleArgs exposed + the same frame-side read and gate) — comes out ≈ 720 added for ≈ 710 deleted: not a step either.
• The 500-case harness campaign was not re-run: no source changed on this head (frames: residue 1 — D list, FRAME_HAVE_* off the client entry, lean re-ask #3863 ran it clean on the same code).
• Suggested next from the residue list (§5): R.insert (row 4, −196) and S-ref to the table (row 7, ≈ −250), neither of which this finding touches.

[changeset-bot]

⚠️ No Changeset found

Latest commit: f127083

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

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

[ryansolid]

Folded into #3860 (wip/frames-tiers-integration) per the maintainer — "merging is a bottleneck." The stack was linear, so the integration branch was fast-forwarded over this PR's two commits (a47bbe8f7, f12708386) to #3868's head d9395e398. The branch is kept.

This PR changed no source: it carries the c15-staging-atomic (e) pin (the gap #3844 named, green on the push form on the integrated head) and the measurement that stopped the preview pull form (+483 br carrier over a −546 ceiling; net −63; the sketch unsound at the same-address refetch). FrameHost.preview / Frame.preview are kept; the trade is recorded under #3860's Public API changes › Changed and its decision 13.

— Claude via Cursor