[Decision] The `framework` per-chunk eager-closure ceiling leaves 177 bytes for the whole repo — two ruled user-facing fixes cannot land, and no wording of either fits

[os-project-manager]

Filed by the domain:ui execution seat (PM session session_01EMrWaQw3XS5DxTHxp4yRyC). ⛔ Not graded and no domain:* — routing, type and priority are the triage seat's to produce.

This is one governance question, not two card-level problems. Two separately-ruled cards in this lane both need user-facing copy in the eagerly-loaded locale packs, and both meet the same ceiling. Deciding it once is cheaper and more coherent than letting each card discover it.

The blocking fact

PER_CHUNK_GZIP_CEILINGS['framework'] in scripts/check-eager-closure-budget.mjs is 524,000 bytes gzipped (511.7 KB). On main alone the chunk measures 523,823 bytes — 177 bytes of headroom for the entire repository.

⭐ The load-bearing reading is not that the footnote is too big. It is that the mechanism does not fit with every string deleted:

build	framework gzip	vs. ceiling
main alone (control, worktree detached to origin/main)	523,823 B	177 B under
PR #7391 as delivered	524,476 B	over by 476 B
PR #7391 with the ten-pack copy deleted entirely	524,053 B	still over by 53 B

• The two i18n keys across ten locale packs cost 423 B.
• The mechanism's code — one constant, one helper, one note component, zero strings — costs 230 B. That alone exceeds the 177 B available.

~600 B were already recovered before this was escalated (copy tightened to the ruling's own wording, the fallback default read from the en pack instead of retyped, two test-only data- attributes dropped) and were kept on their own merits. There is no further shaving that closes a 53 B gap on a 230 B mechanism against 177 B of room.

Two independent measurements agree

1. The lane's own control build, base a6e5f7050, worktree detached to origin/main — the table above.
2. CI, read directly from the Bundle Analysis job log on PR feat(non-grid): a platform row ceiling with a loud footnote, and the settled-schema convergence #7391's head 91876eb55 (2026-09-02T17:40:21Z), which I read myself rather than taking from the PR comment:

✅ Console eager closure is 3177.8 KB gzipped across 48 of 516 chunks (budget: 3191.4 KB, headroom: 13.6 KB).
❌ 1 eager chunk is over its per-chunk budget:
 ✅ vendor-objectstack 926.1 KB / 944.3 KB ceiling (headroom 18.2 KB)
 ❌ framework 512.2 KB / 511.7 KB ceiling (OVER by 0.5 KB)
 ✅ ui-components 387.3 KB / 389.6 KB ceiling (headroom 2.3 KB)
✅ Ceiling freshness: all 3 ceiling constants are unchanged on `main` since this checkout's base

⇒ The aggregate ceiling passes, the sensitivity verdict passes, the freshness verdict passes. Only the per-chunk verdict fails. This is exactly the case #5490 built the per-chunk half to catch, and it is working as designed — ⛔ this card is not a claim that the gate is wrong.

The trend, from two directly-comparable readings of the same checker

The R39 seat posted a full-build reading on 2026-08-31 (e33b44796, pnpm check:eager-closure, exit 0) on #6631. Against today:

	2026-08-31	2026-09-02 17:40Z	move
framework measured	492.9 KB	512.2 KB	+19.3 KB
framework ceiling	500.0 KB	511.7 KB	+11.7 KB
framework headroom	7.1 KB	−0.5 KB	exhausted
aggregate measured	3149.2 KB	3177.8 KB	+28.6 KB
aggregate ceiling	3191.4 KB	3191.4 KB	unchanged
aggregate headroom	42.2 KB	13.6 KB	−68%

⚠️ Two things worth naming, neither of which I assert beyond what these rows support:

• The aggregate ceiling has been ratcheting down over time (4,086,000 → 3,345,000 → 3,300,000 → 3,268,000, enumerated on finding: the eager-closure headroom ruling A was decided on (2.02% / ~78 KB) is now 0.905% / 29.3 KB — and I could not establish which side moved #6631) — the ruled direction of travel. But the framework per-chunk ceiling has been moving up (500.0 → 511.7 KB in two days), and its payload has grown ~1.65x faster than the ceiling was raised. Those are opposite directions.
• The headroom is being consumed in real time by ordinary lane work. The implementing dev observed framework move 510.8 → 511.5 KB on main during the ~90 minutes of its own measurements. ⚠️ Disclosure: part of that was mine — landings I sequenced this round (core: ValueDataSource.matchesASTFilter applies NO filter to a flat implicit-AND array or to is_null / is_not_null — every row comes back, silently (measured under #7221) #7349, finding(components/plugin-detail): the action-id → ActionDef lookup now exists twice, and the two copies disagree about mixed arrays #7182, finding(cli): validate / check print only TOP-LEVEL zod issues, so every union-shaped schema reports a bare "Invalid input" and the per-arm guidance is never seen #7004) went into main inside that window. Treating eager-closure bytes as a shared serial resource, rather than as a per-file surface, is a miss I have already recorded publicly against this seat. It bears on this card because it means "wait for headroom to appear" is not a stable plan: absent a decision, the next lane to measure gets the room.

Who is blocked right now

card	ruling	state
#7210 + #7225 (PR #7391)	ruling a′ requires a role=note footnote naming both N and M	delivered, complete, check:eager-closure RED — parked
#7194	ruled refusal copy lands in the same ten packs	in flight; warned, expected to meet the same 177 B

⇒ #6631's closing comment (2026-09-02T01:48Z) deliberately left this exact question unfiled, and pre-registered the condition for filing it:

The domain:ui seat's residual policy question … is a gate-policy choice with no measured pull today … It stays recorded here; reopening is free if a chunk ceiling is hit by drift rather than by a change. Not filed as a card (no consumer is waiting on it).

Both stated conditions now hold: there is a measured pull, and there are two consumers waiting. ⛔ I am not reopening #6631 — its own question ("which side moved") was answered correctly and it is rightly closed. This is the deferred half, filed as its own card as that comment anticipated.

What is NOT being asked

⛔ The lane did not raise the ceiling and did not weaken the ruled footnote below what ruling a′ specifies. Both were available and both were refused: widening a gate threshold is 门禁削弱 and sits on the human floor, and a ruled fix delivered smaller than its ruling is not a delivery. The fork was reported instead, which I accept as a legitimate delivery, not a failure.

The options

A. Re-baseline framework — the gate's own documented route for intended growth: raise PER_CHUNK_GZIP_CEILINGS['framework'], move PER_CHUNK_BASELINE with it, justify the bytes in the PR.
Cost: 653 B of eager payload, and it spends a governance ratchet. Buys: the ruled safety footnote in ten languages across four views, and unblocks #7194 on the same decision.

B. Shrink the eager payload first, then land both unchanged. packages/i18n's entry statically re-exports all ten locale packs, which is why they are eagerly reachable at all; making that lazy frees far more than either card costs.
Cost: an architectural change to a published entry — a separate ruled card (#5324 / #6795 already name candidates), out of scope under the current dispatches, and it blocks both PRs until it lands.

C. Take #7148's precedent literally. Ruling a′ names that chart footnote as the precedent "for placement and tone", and that footnote is plain untranslated English in JSX with no i18n keys.
Cost: the footnote stops being translated in ten locales, and it still does not fit (the 230 B code row) — so C reduces the ask rather than closing it.

⚠️ A considered non-option, recorded so it is visibly rejected rather than overlooked: routing the locale packs into their own chunk via advancedChunks would make the per-chunk verdict green (the aggregate has 13.6 KB of room) without removing a single byte from what the browser fetches. That is A with the cost hidden — it grants headroom administratively while the gate's own text warns against widening "just to get a green check". ⛔ Not proposed.

四维分析

长期健全性(权重最高,本卡由它决定) —— 上限是载荷健康的代理指标,不是健康本身。一个只剩 177 字节的阈值,已经无法拒绝任何有意义的回归(它要拦的是 89 KB 那一类),却足以拒绝一个 230 字节的正确性修复 —— 此时它不再是在治理增长,而是在按谁最后测量任意配给。⚠️ 但反向同样成立:framework 两天内上限升 11.7 KB、载荷升 19.3 KB,若每次撞墙都选 A,棘轮就是单向松开的,而聚合侧刻意向下棘轮的设计意图会被逐块架空。⇒ B 是唯一同时修复代理指标与被代理事物的选项;A 保住修复但欠下一笔必须显式记账的债。

真实业务需求 —— 已测量,非推测。#7210 来自真实应用:页脚产生了一个假结论,需要一次浏览器会话才能证伪;静默截断是同一类问题下沉一层。ruled a′ 的脚注是区分"完整排程"与"被截断排程"的唯一在屏信号。#7194 的拒绝文案同理。⇒ 需求侧支持尽快落地,即 A 或 C。

让 AI 写的代码不易出错 —— 这条明确反对 C。删掉或不翻译脚注,恰好恢复 ruled a′ 存在就是为了禁止的那种沉默,且是对十个语言的用户恢复。⚠️ 更值得记的是:C 连自己的目标都达不到 —— 230 字节的机制码仍然超 53 字节,所以 C 是"付出全部代价、只换来一个更小的请求"。⇒ 反对 C。

创业公司范围纪律 —— 唯一支持 C 的轴,且不足以压过其余三条:文案只有两个键、已是裁决自身形式允许的最紧措辞、两种做法都不新增可编写面。⚠️ 本轴真正的分量在别处:它反对 B 作为前置条件 —— 让一个已裁定的安全修复去等一次未排期的、对已发布入口的架构重构,是这类修复"悄悄永远不落地"的标准路径。⇒ 若选 B,建议同时裁定 #7391 与 #7194 的处置(等待、或先落机制后补文案)。

实施者建议:A,并把 B 作为后续卡立即立单。⛔ 本席位不代裁 —— 门禁阈值削弱与已发布入口的架构变更都在人工地板上,两者皆非执行席可决。

Related

#6631 (this card's provenance — its closing comment pre-registered this filing) · #5468 (ruling A/C/B on the eager-closure budget) · #5490 (why the per-chunk half exists) · #5924 / #6683 / #6776 / #6778 (the aggregate re-baseline chain and its provenance repair) · #5324 / #6795 (named payload-shrink candidates) · #5266 (the 89 KiB regression class the gate exists to catch) · #7148 (the untranslated-footnote precedent ruling a′ cites) · #7210 / #7225 / PR #7391 · #7194

[os-project-manager]

Second lane has now hit the same ceiling — and reading the gate's own source changes what this decision is about

domain:ui execution seat, PM session session_01EMrWaQw3XS5DxTHxp4yRyC. Three things arrived after filing: the second fork is now measured rather than predicted, the two forks turn out to have different shapes, and the gate's own provenance comments record two prior re-pins of this exact ceiling for this exact reason. All figures below are read from scripts/check-eager-closure-budget.mjs on origin/main directly, not relayed.

1. The second fork is real — #7194 / PR #7400

Delivered, draft, red on check:eager-closure only. Measured against a control build of the base, in two halves:

build	framework gzip	delta	gate
control (a6e5f7050)	523,823 B	—	exit 0 (headroom 177 B)
mechanism only (renderer + tests, packs at base)	523,823 B	0	exit 0
mechanism + ten-pack copy (eacc55913)	524,243 B	+420	exit 1, over by 243 B

2. ⭐ The two forks are NOT the same shape, and no single option resolves both

	mechanism bytes	copy bytes	fits with copy deleted?
#7210 / PR #7391	230	423	⛔ No — still 53 B over
#7194 / PR #7400	0	420	✅ Yes — gate exits 0

⇒ Option C (ship untranslated English, defer the ten packs) resolves #7194 and does NOT resolve #7391. When I filed this card I had one fork and presented C as a single option; that was accurate for #7391 and is wrong as a general answer. Only A or B resolves both.

3. ⭐⭐ The gate's own header records this ceiling being raised TWICE, for this exact mechanism

Verified by reading the source, not inferred:

re-pin	framework ceiling	what the bytes bought
objectui#6759	502,000 → 512,000	"user-facing diagnostic text, in ten locales, for two ordinary author mistakes that used to be unreported"
objectui#7173	512,000 → 524,000 (+4,671 B)	38 keys in ten languages for the AI approvals inbox

Both comments make the same observation the two blocked cards make now. On #6759:

The CODE that fix added buys none of this chunk, which is why there was nothing to trim instead: plugin-timeline is lazy and is not in the eager closure at all, so the guard, the value speller and both refusal branches cost this budget zero. Every byte of the growth is packages/i18n locale data, which lands here.

And on #7173, explicitly naming it as a recurrence:

The COMPONENT costs this budget nothing: plugin-chatbot is lazy and is not in the eager closure. Every byte is packages/i18n locale data, which lands in this chunk — the same mechanism as objectui#6759's twenty strings above.

⇒ #7400 is the third instance of a pattern the gate has documented twice: a lazy plugin's diagnostic costs the eager closure nothing, and its translations cost all of it. ⚠️ I record this neutrally. It is genuine precedent for A — and "this ceiling has now been raised three times in a fortnight for translations" is equally a reason to ask whether B is overdue. Both readings live in the same fact.

4. ⭐⭐⭐ The decisive reading: neither blocked card caused this — 98% of the headroom went to drift

PER_CHUNK_BASELINE['framework'] is 514,863, pinned at #7173's measured head. The ceiling is 524,000.

	bytes
headroom the gate was designed to carry at the last re-pin	9,137 (0.10x the 91,136 B regression)
main today	523,823
⇒ drift since the baseline was pinned	8,960
⇒ headroom left	177 (0.002x)

Drift has consumed 98.1% of the headroom this ceiling was deliberately given. Against that, the two blocked cards ask for 1,073 bytes combined (653 + 420) — comfortably inside the 9,137 the gate intended to carry, and roughly one eighth of what drift took while nobody was measuring.

⚠️ Both prior re-pins were made under the same condition, and both said so — #6759 at 1.8 KB left (0.02x), #7173 at 1,808 B left (0.02x). Today is 177 bytes, 0.002x — an order of magnitude tighter than either occasion that triggered a re-pin. #6759's comment names where this belongs:

The drift itself, and that nobody could say which side moved it, is objectui#6631; this card does not close it and deliberately leaves the other two entries alone.

⇒ This is the third card in a row to pay for drift it did not cause, and the second to be blocked outright by it.

5. If option A is chosen, what it mechanically requires

Not a recommendation — the arithmetic, so the choice is costed. The gate enforces two constraints, both checked in scripts/__tests__/check-eager-closure-budget.test.ts: every ceiling must pass on the measurement in PER_CHUNK_BASELINE, and its headroom must stay under REGRESSION_THIS_GATE_MUST_CATCH_BYTES. Re-pinning onto today's main plus both cards (523,823 + 1,073 = 524,896) at the file's own ~0.10x convention gives a ceiling near 534,000, with the baseline moved to the fresh measurement in the same commit. ⚠️ That is a ~10 KB move, not a ~1 KB one — because it re-pins the gauge onto reality rather than only buying the two cards their bytes. Choosing A while moving the ceiling by only the 1,073 bytes the cards need would leave the gauge exactly as stale, and the next card would be back here.

⛔ The gate also forbids the opposite move — "Do not LOWER one below the measured figure to express an aspiration: this is a ratchet, and a ceiling under today's reality is a gate that lands red on main" — so option B's payload shrink must actually land before either ceiling comes down.

Status of the two blocked cards

Both parked, neither weakened, neither ceiling touched. #7210 / #7225 → Blocked-by: #7399, Unlock-action: re-check PR #7391. #7194 → Blocked-by: #7399, Unlock-action: re-check PR #7400. Standing-down comments on both PRs. #7400's Clause-② contract review is running in parallel at CONTRACT_REVIEW_TIER, so it is ready the moment this rules.

⚠️ Also surfaced by #7400's implementer and filed separately as #7401: a registration's defaultProps.chartType is inert on the SDUI path, so scatter-chart, plugin-charts:scatter-chart and pie-chart all render as two-series bar charts — which also means widget-dom-leak-sweep's pie / donut / radar / scatter entries have been passing as bar charts. Unrelated to this decision; noted so it is not lost in it.

---

Generated by Claude Code

[os-project-manager]

Measurement refreshed against current main — the gap is byte-identical 7 hours and ~9 commits later

domain:ui execution seat, PM session session_01EMrWaQw3XS5DxTHxp4yRyC. ⛔ Not a re-ask. Both blocked PRs were carrying readings taken at 17:40Z and 18:25Z against bases that are now long superseded, and this card's whole argument is quantitative — so the numbers are refreshed rather than left to be argued from stale data. The options and the ask are unchanged and are not restated here.

What I did

Updated both PR branches onto current main (a merge commit each — ⛔ no rebase, no force-push, so neither dev's checkout is invalidated) and let CI re-measure.

	old base	new base
PR #7391 (#7210 + #7225)	a6e5f7050	bf244f400
PR #7400 (#7194)	0614b6df1	bf244f400

bf244f400 is main including tonight's landings — #7212 (6aeba673c) and #7204 + #7224 (bf244f400 itself), among ~9 commits.

The fresh reading, from PR #7391's job log

✅ Console eager closure is 3178.9 KB gzipped across 48 of 516 chunks (budget: 3191.4 KB, headroom: 12.5 KB).
❌ 1 eager chunk is over its per-chunk budget:
 ✅ vendor-objectstack 926.1 KB / 944.3 KB ceiling (headroom 18.2 KB)
 ❌ framework 512.2 KB / 511.7 KB ceiling (OVER by 0.5 KB)
 ✅ ui-components 387.3 KB / 389.6 KB ceiling (headroom 2.3 KB)
✅ Ceiling sensitivity … ✅ Ceiling freshness: all 3 ceiling constants are unchanged
 on `main` since this checkout's base (bf244f400…)

⭐ framework 512.2 / 511.7, OVER by 0.5 KB — byte-identical to the 17:40Z reading. PR #7400 is red on the same per-chunk check on its own refreshed head (15ad6a762); I have not re-read its exact figure and am not quoting one I did not measure.

Why the refresh is worth having rather than noise

Two things it establishes that the stale reading could not:

1. The rebuild is real, not a cached artifact. The ui-components chunk hash moved (ui-components-Bd4e_cV4.js → ui-components-DHsni2p8.js), so main genuinely changed underneath and the closure was re-weighed — and still weighs the same for framework. A delta of zero from an unchanged build would prove nothing; this one has a control.
2. The ceilings have not been touched. Freshness passes against bf244f400, and PER_CHUNK_GZIP_CEILINGS['framework'] is still 524_000 with PER_CHUNK_BASELINE still 514_863 on current main. So nothing has been re-baselined out from under this card, and nothing has shrunk the payload.

⇒ The situation is not drifting toward resolution on its own. main's headroom has stayed at roughly 200 bytes all night while nine commits landed, which is the same fact this card was filed on — now measured twice, seven hours apart, either side of three of this lane's own landings.

⚠️ For completeness on the other side of the ledger: all three cards this lane dispatched tonight were chosen for byte-neutrality and all three measured clean — #7212 (0 files added/removed across 33 packages, Bundle Analysis green), #7204 + #7224 (eager closure delta 0 across all 48 chunks, with a differing entry-chunk hash as the control), #7115 (docs and scripts/ only). None of them consumed a byte of the headroom this card is about.

---

Generated by Claude Code