Uh oh!
There was an error while loading. Please reload this page.
perf(app-shell): take metadata-admin out of the console eager closure, and re-baseline the ceiling with it - #6913
Conversation
`AppContent` declares six `lazy()` imports of `views/metadata-admin/index.ts` and none of them deferred anything. Two static edges held the module, and the first one is what made the other unremovable: 1. the module ran FIVE registrations at load, so `@object-ui/app-shell`'s published `sideEffects` array named it (#6683) and no bundler could shake it; and 2. the package barrel re-exported 25 runtime values FROM it, which the console's entry imports statically, and `services/builtinComponents.tsx` held `MetadataDirectoryPage` / `MetadataResourceRouter` by value. So every page, preview and inspector under `views/metadata-admin/` was fetched and parsed before first render — 172,945 gzipped bytes, 144 modules, the largest entry the declared-lazy ledger ever carried. Split the two concerns instead of cutting either edge where it was observed: - the five registrations move to `views/metadata-admin/register-builtins.ts`, bare-imported by the PACKAGE ENTRY, so they stay exactly as eager as they were and the page barrel becomes a pure re-export module. The `sideEffects` array names the new leaf. - the barrel's 25 runtime re-exports (and 11 type-only ones, which are erased and carry no edge) name the LEAF modules. Same names, same types. - `metadata:directory` / `metadata:resource` register `lazy()` values, each with its own `Suspense` inside the registration value — the shape the already-lazy registrations in `apps/console` use, so `registerAppComponent`'s signature is untouched and no render site changes. Measured from `apps/console/dist/eager-closure.json` on two full builds: 3,254,230 -> 3,222,314 gzipped bytes (-31,916 B); the `metadata-admin` chunk leaves the eager set entirely (172,945 -> 0) and 45 -> 48 eager chunks. The ledger line for `views/metadata-admin/index.ts` is deleted — its `missing` half fails the build until it goes, which is how a win here gets recorded. The trap is written up beside `builtinComponents.tsx`: doing ONLY the `lazy()` there is worth -30 B, GROWS the chunk, leaves it EAGER, and passes every gate, because `ineffective-dynamic-import` cannot see a static edge that lives in another module. That is the #5486 shape. Refs #6776 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
payload 3,300,000 -> 3,268,000 over a measured 3,222,314 on `3d257c85a`, with {@link BASELINE} moved in the same commit so the two checked constraints hold. Taking 31,916 bytes out without moving the ceiling leaves 75.9 KB of headroom — 0.85x the 89 KiB regression this gate exists to catch, which the gate itself printed on the post-change build. That is the blind band reopening the day the win lands, and it makes the optimisation net-negative: the budget table stops being able to see the next regression. The maintainer ruling of 2026-08-30 on #6776 wrote this down rather than leaving it to the implementer, verbatim: ⛔ ceiling 处置写死(不作实施者临场判断):-31KB 把余量推到 ~0.89x 门禁 89KB 回归阈值(近盲),同批重设 `MAX_EAGER_CLOSURE_GZIP_BYTES`;抬 ceiling 是有申报 程序的 ratchet,裁决原话引入 PR 正文。 Direction is DOWN: no build that passed before this edit and measures under 3,268,000 fails after it. REGRESSION_THIS_GATE_MUST_CATCH_BYTES is untouched, and the three per-chunk ceilings are deliberately left alone — this change moved none of those chunks. Refs #6776 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
…e predicted ones The ruling that ordered this comment predicted the in-fence-only change (making the value lazy in `builtinComponents.tsx` alone) at -30 B on the eager closure and +189 B on the `metadata-admin` chunk. Measured it rather than transcribing it: a full console build of `fab4802e3` carrying ONLY that change, against the same commit unmodified. eager closure 3,254,230 -> 3,254,441 B gzipped (+211 B) `metadata-admin` chunk 172,945 -> 173,341 B gzipped (+396 B) eager chunk count 45 -> 45 of 513 (UNCHANGED) Both directions of the qualitative claim reproduced — the chunk grows, it stays EAGER, and every gate is green (build exit 0, `ineffective-dynamic-import` prints its usual 43 pinned entries with no 44th, `declared-lazy-views` prints "2 eager, all pinned"). The closure figure did not: it came back POSITIVE, so the change costs bytes in both places rather than buying 30. Both write-ups now carry the measurement and say the prediction differed. Refs #6776 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
…-baseline Two assertions in `check-eager-closure-budget.test.ts` carry `BASELINE.gzipBytes` as a RENDERED literal — deliberately, so a rendering derived in the test cannot agree with the renderer by construction and pin nothing. Moving the baseline to 3,222,314 moves that rendering from 3177.7 to 3146.8, so both literals move with it. The comment now names both re-baselines rather than only #6683's. Refs #6776 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
os-sam
commented
Aug 30, 2026
Clause-② contract review — verdict: ACCEPT WITH FOLLOW-UP
The one follow-up is non-blocking and stated at the bottom. Everything the ruling ordered is in the diff, and everything the PR body claims that I re-measured came back exactly as claimed. 1. Published surface — verified against the emitted |
base fab4802e3 (my build) | head ae9f949ac (my build) | PR claim | |
|---|---|---|---|
| eager closure gz | 3,254,230 | 3,222,314 | matches to the byte |
| eager / total chunks | 45 / 513 | 48 / 517 | matches |
metadata-admin | 172,945 B gz, in files[], 144 modules | absent from files[], 26 modules, 35,339 B gz on disk | matches |
| Δ | −31,916 B | matches |
The zero has its controls: the same files[] query hits MetadataInspector and runtime-metadata-persistence, and metadata-admin-C5OJvVve.js exists on disk — lazy, not deleted. eagerGzipBytes equals the sum over files[]. A post-review rebuild of head reproduced 3,222,314 byte-for-byte. The −31,794 ruling projection vs −31,916 measured (122 B, bases differ) is accepted as reported-not-reconciled per the dispatch.
4. The E2 trap — direction and green-gates claim independently reproduced; one robustness note
I rebuilt fab4802e3 carrying only the head version of builtinComponents.tsx (nothing else):
- chunk still EAGER, 45/513 unchanged,
metadata-adminchunk +183 B (173,128), build exit 0,ineffective-dynamic-importprinted its usual 43 pinned entries with no 44th,declared-lazy-viewsprinted "10/12 … 2 eager, all pinned". Every gate green through a change that defers nothing — the claim the trap note is for, reproduced in full. - The closure delta in my run was −7 B, not the dev's +211 B. My edit is not byte-identical to the dev's E2 experiment (theirs predates the final file text), and ±~200 B on a 3.25 MB closure is evidently within the variance the byte-form of the edit induces — note my chunk delta (+183) lands near the ruling's +189 prediction while the dev's was +396. So the robust invariant is: an in-fence-only lazy() pays nothing beyond noise in either direction, grows the chunk, leaves it eager, and no gate can see it. The specific "+211 B, costs bytes in both places" is true of the dev's run but its sign does not survive re-derivation from a different byte-form of the same change.
5. The ceiling — arithmetic and rule both check
REGRESSION_THIS_GATE_MUST_CATCH_BYTES = 89*1024 = 91,136. New headroom 3,268,000 − 3,222,314 = 45,686 B = 0.5013x (gate prints "headroom 44.6 KB = 0.50x the 89.0 KB regression", exit 0). The file documents H = REGRESSION / 2 explicitly, and the two prior re-baselines sat at 0.5047x (#6683) and 0.4949x (#5924) — the discipline is real and matched. Had the ceiling stayed at 3,300,000: 77,686 B = 0.852x — the PR's "0.85 x" is the measurement, the ruling's 0.89x was an estimate, and the PR says so. Direction is down; the ruling's ratchet caution concerns raising. REGRESSION_… and the three per-chunk ceilings are untouched, as stated.
6. The ablation — re-run, not just audited
Re-pinned packages/app-shell/src/views/metadata-admin/index.ts in DECLARED_LAZY_VIEWS_STILL_EAGER on the fixed tree myself: blob 5cc79a16… → eaa76737… (both hashes match the PR's stated pair exactly — my mutation was byte-identical), rebuild failed exit 1 with the verbatim RolldownError: [declared-lazy-views] 1 pinned view(s) are NO LONGER in the eager closure, restore verified by hash and clean git status. The error fires because the view left the eager closure — the deleted ledger line is a demonstrated win. The restored head tree rebuilds green with the plugin printing "11/12 … 1 eager, all pinned" (moved from 10/12 / 2 eager, as claimed).
7. The known failure is genuinely pre-existing (#6893) — checked, not waved through
Because this PR edits the array, this was treated as a live question: (a) on head with dist built, check-sdui-registration-pins.test.ts fails 1/11, exactly the mcp:connect-agent src-vs-dist spelling; (b) control: the same single failure fires on unmodified base fab4802e3 with dist built; (c) with the dist artefact moved aside, head passes 11/11 (restored after); (d) the diff swaps only the metadata-admin pair within each spelling block — the ConnectAgentWidget entries and the dist-before-src walk order derivePinnedKeys depends on are untouched. Pre-existing, correctly attributed to #6893.
8. The changeset — every figure re-verified; correct
3,254,230 → 3,222,314 / −31,916 B, 172,945 B chunk leaving the eager set, 25 runtime + 11 type-only, root-only exports, six lazy() declarations (AppContent.tsx:63-68), 3,300,000 → 3,268,000, 0.85x — all re-measured above and all correct. Patch bump for a shape-preserving change is consistent with the identical d.ts surface; Changeset Bump Policy / Fixed Group / Declaration checks are all green.
Verified vs taken on report
Re-ran/measured here: both console builds and artefacts; the d.ts surface diff; the three gates (side-effects-array + two mutation controls, sdui-registration-pins, eager-closure); the E2 reproduction; the ablation; ceiling constants and arithmetic; the #6893 controls; gate test files vite-declared-lazy-views / check-eager-closure-budget / check-side-effects-array / side-effects-declaration-consistency — 162/162; app-shell targeted slice views/metadata-admin/ + services/ — 228 files, 2391 passed / 1 skipped; type-check exit 0; the app-shell diff 3d257c85a..head is comment-only (confirms the suite basis).
NOT MEASURED here: the full pnpm exec vitest run packages/app-shell/ locally — my run hit its 560 s budget, exit 143, which is no reading. Covered instead by CI's four test shards (all green on the merge ref) plus the targeted slice above; no conclusion in this review depends on the local full-suite run. The dev's 577-file/5674-pass figure is taken on report + CI. check:readme-exports, check:entry-guard, check:self-import, check:phantom-deps, check:control-bytes, check:docs-route-closure, i18n gates, and lint were taken from CI (all green), not re-run locally. The dev's E2 figures (+211/+396) are their measured run, taken on report — my independent variant measured −7/+183 (see §4).
CI at review time: all 30 check runs completed — 27 success, 3 skipped (coverage variants, dependabot), zero failures. Test shards 1–4, Type Check, Lint, Build & E2E, Bundle Analysis, Governed Surface Queue Guard: success.
Follow-up (concrete, non-blocking, fixable on this branch)
- Make the trap note robust to the sign flip it will meet.
builtinComponents.tsxand thevite-declared-lazy-views.tsheader both state the E2 closure cost as "+211 B … it costs bytes in both places". An independent E2 re-run (this review, head's exact file text onfab4802e3) got −7 B closure / +183 B chunk — same verdict, different sign on the closure half. A next person who tries their own variant, measures a small negative number, and concludes the note is wrong is exactly the reader this note exists for. Suggested one-line addition to both comments: the closure delta's sign varies with the byte-form of the edit (runs have measured +211 and −7); the invariant is that it never pays more than noise, the chunk grows, the chunk stays EAGER, and every gate stays green. Keep the dev's measured figures and the ruling's prediction as they stand — this adds the third data point rather than replacing the two.
Landing
This PR stays draft and is not mine to close, mark ready, or merge — and not the dev seat's either. Per the ruling's designation of the sideEffects array as a governed surface: "⛔ 治理面(#6683 sideEffects 数组)PR 走人工合并道,不 arm auto-merge."Human merge path only; auto-merge is not armed and must not be. Draft status and the green Governed Surface Queue Guard are consistent with that at review time.
Generated by Claude Code
os-sam
commented
Aug 30, 2026
PM: clause ② verdict received — ACCEPT WITH FOLLOW-UP. And the follow-up corrects a call of mine.
Both landing constraints are confirmed intact by the review: clause ② was reviewed by content (the
|
| E2-only, what reproduced | dev | reviewer |
|---|---|---|
| chunk still EAGER | yes | yes |
| eager chunks 45 / 513 unchanged | yes | yes |
| build exit 0, every gate green | yes | yes |
ineffective-dynamic-import 43 pinned, no 44th | yes | yes |
declared-lazy-views "2 eager, all pinned" | yes | yes |
| closure delta | +211 B | −7 B |
The lesson the trap comment must carry is "the closure barely moves and the module stays eager,
because the static edge that defeats the import() lives in another module" — not a signed byte
figure. A comment that presents ±200 B as the finding invites the next reader to re-derive the sign
and conclude the comment is wrong. That is the same defect class this seat has been closing all
round: a stated measurement outliving the conditions it was taken under.
⇒ I am treating the reviewer's follow-up as required before a human merges, not optional. It is
comment text only and does not touch the change.
What the review established, that I am relying on
Reproduced to the byte rather than taken on report:
- Published surface — all 233 exports of the emitted
dist/index.d.tsenumerated with the
compiler API on both trees, byte-identical after path normalization. 25 runtime + 11 type-only
confirmed;registerAppComponent's signature unchanged;exportsroot-only; zero deep imports
with the control hitting.dist/index.js:143bare-imports the leaf with all five calls, and the
emitted page barrel has 0 — against a base control that hit. - The governed half — ⭐ the
sideEffectsarray is proven correct, not merely green: both
mutation controls fire. Droppingregister-builtins(MISSING) and re-addingmetadata-admin/index(STALE) each exit 1 naming both spellings. A gate that only ever prints ✅
has not been shown to discriminate; this one has. - The headline measurement — both consoles rebuilt from scratch, every figure matching to the
byte: base 3,254,230 / 45 of 513 / 172,945 gz EAGER / 144 modules; head 3,222,314 / 48 of 517 /
absent fromfiles[]/ 26 modules; Δ −31,916 B. Zeroes controlled, head rebuild deterministic. - Ceiling — 45,686 B = 0.5013x of 91,136, against the file's own documented
H = REGRESSION / 2
rule and priors of 0.5047x / 0.4949x. The old ceiling would have carried 0.852x. Direction down. - Ablation — re-run by the reviewer: blob pair
5cc79a16…→eaa76737…matches the dev's exactly,
theRolldownErrorverbatim on exit 1, restore hash-verified. - finding(test-infra):
check-sdui-registration-pins.test.tsfails on any tree wherepackages/app-shell/distexists #6893 — the 1/11 failure reproduces identically on unmodified base withdistbuilt, and
goes 11/11 withdistaside. The edit never touches themcp:connect-agententries or the walk
order ⇒ genuinely pre-existing, and ⛔ not this PR's to fix.
NOT MEASURED, declared: the full local app-shell suite (exit 143 at the reviewer's 560 s budget —
no reading, neither red nor green). Covered by CI's four green test shards plus a 228-file targeted
slice at 2391 passed. No conclusion above depends on it.
CI at review time: all 30 checks completed — 27 success, 3 skipped, zero failures.
Next
Dispatching the follow-up now: a robustness line on the two trap comments so neither presents a
sign-unstable byte delta as the finding. Comment text only, on this branch. The PR stays draft,
auto-merge stays unarmed, and it remains a human merge.
Generated by Claude Code
… notes The E2 trap (making the value lazy in `builtinComponents.tsx` alone) is real and its write-up stays. What does not stand is the SIGN of the closure delta it quotes. A second, independent rebuild of the same variant measured that delta at -7 B where the recorded run measured +211 B. Both figures stand as what their run measured; together they say the delta is small and sensitive to the byte-form of the edit, so its sign carries no finding. Everything else reproduced identically on both rebuilds: the chunk stays EAGER, the eager chunk count stays 45 of 513, the build exits 0, `declared-lazy-views` prints "2 eager, all pinned" and `ineffective-dynamic-import` prints its usual 43 with no 44th. That pair -- the EAGER verdict and the unmoved 45/513 -- is the finding, and both notes now say so, so the next reader who re-derives the byte figure and gets the other sign does not conclude the note is wrong. Comment text only. No recorded measurement is deleted; the +211 B and +396 B figures stay exactly where they were. Proven comment-only two ways: all 16 changed lines are added lines inside an existing JSDoc block (0 removed), and each file's non-comment body is byte-identical to its pre-edit blob (`builtinComponents.tsx` 5,629 B, sha256 cc0fac46...; `vite-declared-lazy-views.ts` 20,977 B, sha256 c2802841...). Refs #6776 Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
os-sam
commented
Aug 30, 2026
Follow-up commit What changedBoth E2 trap notes now say that the SIGN of the closure delta is not the finding:
Each note keeps the figure it already carried (+211 B closure, +396 B chunk — no recorded measurement was deleted) and adds that a second, INDEPENDENT rebuild of the same E2 variant measured that closure delta at −7 B. Both figures stand as what their run measured; together they say only that this delta is small and sensitive to the byte-form of the edit. What both rebuilds reproduced identically is now stated as the finding: the chunk stays EAGER and the eager chunk count stays 45 of 513 (build exit 0, This is documentation robustness, not a repair. Nothing here was red before the commit and nothing claims it was — there is no failing state a comment edit could start from. Comment-only proof, two independent ways1. Line level. 2. Byte level. Each file's non-comment body — the file with exactly the scanner's comment ranges deleted — hashed on both sides:
Comment-range counts are unchanged in both files (34 and 22); only comment bytes grew, +596 and +462, matching the file totals exactly. Checks
Lint was narrowed to the two touched files rather than run repo-wide, and the narrowing is declared: No changeset (nothing release-relevant), nothing under Generated by Claude Code Generated by Claude Code |
os-sam
commented
Aug 30, 2026
PM: the review's follow-up is applied at |
✅ Console Performance Budget
The eager closure is every chunk the entry reaches through static imports — what the browser fetches and parses before the app renders. The entry chunk on its own is a small fraction of it. 📦 Bundle Size Report
Size Limits
|
os-sam
commented
Aug 30, 2026
The re-baselined ceiling holds after |
at ae9f949ac (reviewer-verified) | at 17f7b0784 (bot, this run) | |
|---|---|---|
| headroom | 45,686 B = 44.6 KB | ≈ 44,650 B = 43.6 KB |
| as a fraction of the 91,136 B regression threshold | 0.5013x | ≈ 0.490x |
not exact. The direction is what matters and it is not in doubt: ~1 KB of headroom was consumed by
what landed on main, not by this PR — the follow-up commit was comment-only, with both files'
non-comment bodies proven byte-identical.
⇒ The ceiling is now marginally tighter than the file's own H = REGRESSION / 2 rule rather than
looser. That is the conservative side — less headroom means the gate catches the next regression
sooner, not later — so it needs no adjustment and I am not proposing one. Re-baselining a ceiling
because ordinary base drift nudged it is how a budget stops being a budget.
No action. Recording it because a human merges this PR, and "the ceiling was chosen on a base that
has since moved" is exactly the thing that should be checked before that happens rather than after.
Generated by Claude Code
os-sam
commented
Aug 30, 2026
CI is terminal green on |
| count | |
|---|---|
| total check runs | 30 |
| success | 27 |
| skipped (never ran) | 3 |
| failure / cancelled / timed out | 0 |
| still in progress | 0 |
The three skipped are Test (coverage), Test (coverage shard …) and dependabot — conditional
jobs, NOT MEASURED rather than green, none of them a required gate. All four Test (shard N/4)
succeeded, as did Type Check, Lint, Build & E2E, Bundle Analysis and every doc/changeset gate.
Standing state, so a human merging this does not have to reconstruct it
| axis | status |
|---|---|
| clause ② contract review | ACCEPT WITH FOLLOW-UP — every measurement reproduced to the byte, the governed sideEffects array proven correct by mutation controls rather than by a green tick |
| the review's one follow-up | applied at 17f7b0784, comment-only, non-comment bodies proven byte-identical |
| CI on the current head | terminal green, above |
| eager-closure budget on the current head | PASS — 3147.8 KB against the re-baselined 3191.4 KB ceiling, 48 chunks |
| ceiling vs base drift | headroom ≈ 43.6 KB ≈ 0.490x of the 91,136 B threshold — tighter than the file's own H = REGRESSION / 2 rule after main moved, i.e. the conservative side. No adjustment proposed |
#6893 (the 1/11 check-sdui-registration-pins failure) | pre-existing, reproduced on unmodified base with dist built, already filed. ⛔ Not this PR's |
⛔ Still not armed, and still not mine to land. The ruling designates the sideEffects array a
governed surface for this purpose and routes this PR to a human merge. It stays a draft with
auto-merge unarmed until a human takes it — green, reviewed, and complete is not the same as merged,
and this seat does not close that gap.
One consequence worth naming for whoever picks it up: #6795 (arm C of #6776, the ~141 KB inviews/studio-design/) is deliberately held behind this PR, because both would re-baseline the same
eager-closure ceiling and touch the same sideEffects array. It unblocks when this lands.
Generated by Claude Code
Fixes#6776
Executes the maintainer ruling of 2026-08-30 (comment 5468951901, batch #4, 「同意」): option B — take the 31 KB — plus the ceiling re-baseline in the same batch. Arm C (the ~141 KB in
views/studio-design/, #6795) is untouched.Session: https://claude.ai/code/session_013hfmP9hoMd3dJwTh85J4yB
Result
Measured from
apps/console/dist/eager-closure.json— the artefact, never a source-level search — on two full console builds:metadata-adminchunkfab4802e3(base)ae9f949ac(this PR)The ruling projected −31,794 B; the measured figure is −31,916 B, 122 bytes apart. The base reading was taken on
fab4802e3, which is not the commit the E1b projection was measured on, so a small difference is expected — reported rather than reconciled.The
metadata-adminchunk still exists; it is now reached only through the sixlazy()declarations that always claimed to defer it. Its module list moved too: 144 modules before, 26 after, with the previews/inspectors/anchors/default-schemas half relocating into the eagersrcchunk (where it belongs — those registrations must still run at load), and the page half staying behind the lazy boundary.What changed, and why the edges had to be cut in three places at once
AppContenthas declared sixlazy()imports ofviews/metadata-admin/index.tsfor a long time and none of them deferred anything. Two static edges held it, and the first made the second unremovable:@object-ui/app-shell's publishedsideEffectsarray named it (Should@object-ui/app-shelldeclare asideEffectsfield? Measured at 242.6 KB gzipped — 8x today's console headroom — but it is a published contract whose failure mode is silent #6683) and no bundler could shake it; andservices/builtinComponents.tsxheldMetadataDirectoryPage/MetadataResourceRouterby value.So:
views/metadata-admin/register-builtins.ts, bare-imported by the package entry. They stay exactly as eager as they were; the page barrel stops being side-effectful. ⛔ The bare import must not be attached to the page barrel —bareSideEffectImportinscripts/vite-declared-lazy-views.tsreads that line as "not pure" anddeclared-lazy-viewswould refuse it, which is the same defect under a new name.metadata:directory/metadata:resourceregisterlazy()values, each wrapping itself in its ownSuspenseinside the registration value — the shapeapps/console's already-lazy registrations use.registerAppComponent's signature is unchanged and no render site changes.views/metadata-admin/index.tsis deleted, andMAX_EAGER_CLOSURE_GZIP_BYTESis re-baselined.Re-derived counts (#6785's correction, confirmed): the barrel carries 25 runtime re-exports (each a static edge) and 11 type-only (erased, no edge). Derived mechanically from both trees; the runtime name set and the type name set are byte-identical before and after.
Published surface — clause ② applies by content
Two published contract surfaces move.
1.
packages/app-shell/package.json'ssideEffectsarray (#6683's field). One entry pair swaps:views/metadata-admin/index→views/metadata-admin/register-builtins, in both the./dist/*.jsand./src/*.tsspellings. The five registrations still run at package load.check:side-effects-arrayderives the enumeration mechanically and agrees: "@object-ui/app-shell:sideEffectsnames exactly the 12 module(s) that register at load time, plus its entry forms (27 entries, 434 modules walked)."check:sdui-registration-pinsweighs the built console: "All 14 registration(s) asideEffectsarray promises are present in the built console (517 chunks weighed; the 3 ruled control(s) are in the derived set)."2. The barrel's 25 runtime + 11 type-only re-exports. Same names, same types, different source modules.
What an out-of-package consumer could import before, and after: exactly the same set, by exactly the same specifiers. Verified, not assumed:
exportsis root-only — its keys are"."and"./styles.css", and./styles.cssis a CSS asset, not a JS path. There is notypesVersions. So@object-ui/app-shell/views/metadata-admin/index.jswas never resolvable by any consumer, and is not now.@object-ui/app-shellreturns only@object-ui/app-shell/styles.cssoccurrences. No JS deep import exists anywhere.Which module a name is re-exported from is therefore invisible outside the package. The only observable difference is the one the change is for: a consumer's bundler can now shake the page barrel.
⛔ The ceiling was reset in the same batch — the ruling wrote this down rather than leaving it to me
MAX_EAGER_CLOSURE_GZIP_BYTES3,300,000 → 3,268,000, withBASELINEmoved in the same commit so the two checked constraints hold. Measured, the old ceiling carried 75.9 KB of headroom over the new payload — the gate printed it:headroom 75.9 KB = 0.85x the 89.0 KB regression. That is the blind band reopening the day the win lands, which would make the optimisation net-negative: the budget table stops being able to see the next regression.Ruling, verbatim:
The direction is down: no build that passed before this edit and measures under 3,268,000 fails after it. Headroom is back to
0.50x the 89.0 KB regression.REGRESSION_THIS_GATE_MUST_CATCH_BYTESis untouched, and the three per-chunk ceilings are deliberately left alone — this change moved none of those chunks. (The measured drift was 0.85x rather than the 0.89x the ruling estimated; recorded in the file.)The smallest and most in-fence-looking version of this change is to make the value lazy only in
builtinComponents.tsx. Measured on a full console build offab4802e3carrying only that edit, against the same commit unmodified:metadata-adminchunkBuild exit 0.
ineffective-dynamic-importprinted its usual 43 pinned entries, no 44th;declared-lazy-viewsprinted "2 eager, all pinned". Every gate green, nothing moved, bytes spent on thelazy()/Suspensescaffolding itself. The ledger cannot see it because the static edge that defeats theimport()lives in another module — the package barrel's re-export. That is the #5486 shape.The ruling predicted −30 B and +189 B. The chunk growth, the EAGER verdict and the green gates all reproduced; the closure figure did not — it came back positive. The trap note beside
builtinComponents.tsxand the header ofscripts/vite-declared-lazy-views.tsboth carry the measured numbers and say so.Ablation — the deleted ledger line is a measurement, not an assertion
Re-pinned
packages/app-shell/src/views/metadata-admin/index.tsinDECLARED_LAZY_VIEWS_STILL_EAGERon the fixed tree and rebuilt the console. Prediction: the ledger'smissinghalf fires and the build stops.Mutation proved on disk before the build (HEAD blob
5cc79a16…vs mutatedeaa76737…, anchored line count 0 → 1); restore proved after it (hash back to5cc79a16…,git diff HEADempty for the file). The restore leg rangit checkout HEAD --against the file's ABSOLUTE path, inside a trap onEXIT INT TERM—HEADnamed explicitly so a path-scoped checkout could not take the mutation back out of a polluted index.Counter-probe 1 of the same plugin (every declared view must be found in some chunk) ran green on the fixed build, which is what separates "the view became lazy" from "the matcher stopped matching". The plugin's own line moved from
10/12 modules … genuinely lazy; 2 eager, all pinnedto11/12 … 1 eager, all pinned.Verification
Run against
ae9f949ac, the final commit, unless stated.Builds (2 full console builds for the before/after pair, 1 for E2, 1 for the ablation, 1 final):
pnpm turbo run build --filter='./packages/*'thenpnpm --filter @object-ui/console build— exit 0.Gates — all exit 0 (verdict lines quoted from the gates themselves, exit codes captured before any pipe):
check:eager-closure—✅ Console eager closure is 3146.8 KB gzipped across 48 of 517 chunks (budget: 3191.4 KB, headroom: 44.6 KB)./✅ aggregate closure 3146.8 KB measured / 3191.4 KB ceiling (headroom 44.6 KB = 0.50x the 89.0 KB regression)check:side-effects-array—✅ @object-ui/app-shell: sideEffects names exactly the 12 module(s) that register at load time, plus its entry forms (27 entries, 434 modules walked).check:sdui-registration-pins—✅ All 14 registration(s) a sideEffects array promises are present in the built console (517 chunks weighed; the 3 ruled control(s) are in the derived set).check:readme-exports,check:entry-guard,check:self-import,check:phantom-deps,check:control-bytes,check:docs-route-closure,check:i18n-keys,check:i18n-drift,check-changeset-presence— all exit 0.Tests (from the repo root, per this repo's documented invocation rule):
pnpm exec vitest run packages/app-shell/— 577 files passed, 5674 passed / 1 skipped, exit 0. Run on3d257c85a+1; the app-shell diff from there to head is comment-only (git diff … -- packages/app-shell/has no non-comment+/-line).pnpm exec vitest run scripts/__tests__/{vite-declared-lazy-views,check-eager-closure-budget,check-side-effects-array,side-effects-declaration-consistency,vite-ineffective-dynamic-imports,render-budget-comment,check-docs-route-eager-closure,check-sdui-registration-pins}.test.ts— 255 passed of 256, 1 known unrelated failure (below).pnpm --filter @object-ui/app-shell type-check— exit 0.pnpm --filter @object-ui/app-shell lint— exit 0 (0 errors).pnpm lint:root— exit 0 (0 errors). The 5 new warnings inbuiltinComponents.tsx(3 ×react-refresh/only-export-components, 2 ×no-explicit-any) are the same profile the established precedent carries onmain:apps/console/src/registerDeveloperComponents.tsxhas 7 of exactly these,registerAccountComponents.tsx3.Two pin tests moved with the change, and both moves are stated rather than silent:
vite-declared-lazy-views.test.ts— the "declared side-effectful" pin now namesregister-builtins.ts, and a new pin asserts the page barrel is not declared side-effectful and carries no bare side-effect import. That second one is the assertion that makes the shakeability claim testable without a build.check-eager-closure-budget.test.ts— two assertions carryBASELINE.gzipBytesas a rendered literal on purpose (a rendering derived in the test would agree with the renderer by construction and pin nothing). 3177.7 → 3146.8.Known unrelated failure, already filed:
check-sdui-registration-pins.test.ts› "derives the keys from the arrays…" expects thesrcspelling formcp:connect-agentand gets thedistone on any tree wherepackages/app-shell/disthas been built. Proved independent of this diff by removing that single build artefact and re-running: 11/11 pass, then restored. It is #6893.Landing
⛔ Human merge path. The
sideEffectsarray is #6683's published field and the ruling designates it a governed surface for this purpose: "⛔ 治理面(#6683 sideEffects 数组)PR 走人工合并道,不 arm auto-merge." This PR stays draft: auto-merge is not armed, it is not enqueued, and it is not marked ready. Not by me either, green and reviewed.Generated by Claude Code