perf(components): the eager `ui-components` chunk carries lucide's 1,767-entry dynamic-import map and the whole `icons` record — 45,749 B / 144,361 B gzipped, both measured

[baozhoutao]

Filed bare (no assignee, no domain:*, no pm:*) out of the measurement-first round on objectui#7848, whose fence says a slimming that the census finds "is a second card and a second PR". This is that card. ⛔ Nothing here has been implemented and ⛔ no gate constant was moved.

All figures below are from four full console builds in one container on tip 1f4e02995a, read directly out of apps/console/dist/eager-closure.json. Every mutation was proved on disk (blob-hash before/after plus printed grep anchors) and restored with git checkout HEAD -- path, proved by an empty git diff HEAD.

The reading on 1f4e02995a

ui-components   395,032 gz   ceiling 399,000   headroom 3,968 = 0.0435x the 91,136 B regression
aggregate     3,186,729 gz   ceiling 3,210,000 headroom 23,271 = 0.2554x

ui-components is the only entry in EXHAUSTED_HEADROOM_ALLOWANCES (4,289). The gate itself prints the trip point: reds below 3378 bytes — so the live row has 590 bytes of runway before the ratchet fires. Clearing the 0.10x floor on its own needs the chunk at 389,886 or less, i.e. a saving of 5,146 gzipped bytes.

Where the bytes are

Census of the emitted chunk's own module list (2,240 modules, 2,926,967 B renderedLength):

area	renderedLength	share	modules
lucide-react	1,480,797	50.6%	1,770
packages/components/src/renderers	396,868	13.6%	105
packages/fields/src/widgets	280,718	9.6%	72
packages/components/src/ui	112,206	3.8%	47
packages/components/src/custom	104,106	3.6%	29
react-day-picker	99,431	3.4%	98
react-hook-form	80,343	2.7%	1

Half the chunk is lucide, through two independent doors.

Door 1 — the dynamic-import map is EAGER (dynamicIconImports.mjs, 263,547 B rendered)

Four non-test modules statically import lucide-react/dynamic.mjs:

packages/components/src/lib/lazy-icon.tsx:24
packages/app-shell/src/views/metadata-admin/widgets.tsx:46
packages/app-shell/src/utils/getIcon.ts:16
apps/console/src/utils/getIcon.ts:11

The map is 1,767 () = arrow entries whose only job is to be called later. It does not need to be in the first payload.

Measured (ablation: the four imports stubbed, icons record left alone):

ui-components  395,032 -> 349,283   -45,749
aggregate    3,186,729 -> 3,141,057 -45,672
eager/total chunks  51/528 -> 51/528   (unchanged)

⭐ The two deltas agree within 77 bytes, so this is bytes LEAVING the page load, not bytes moving between columns. No new chunk, no extra request. ui-components headroom becomes 49,717 = 0.55x — above the 0.10x floor and still under 1.00x, so the ceiling stays sensitive rather than going blind. This one change alone is 8.9x the 5,146 bytes the row needs.

Door 2 — the whole 1,767-icon record is EAGER

packages/components/src/renderers/action/resolve-icon.ts:9

import { icons, type LucideIcon } from 'lucide-react';

icons is lucide's runtime record, published as export { index as icons } over a namespace import of a 1,767-line barrel. A namespace object has no dead members, so tree-shaking cannot help: every icon module is statically reachable. Verified on the emitted artifact rather than from source — the path data of air-vent, an icon this product does not use, appears exactly once in assets/ui-components-*.js.

It is also the ONLY non-test module that imports the record (the three other hits are test files), which matches check-lucide-icon-record-names.mjs's declared census of "this file alone".

Measured (ablation: resolveIcon returns null, record import removed):

ui-components  395,032 -> 254,749  -140,283
aggregate    3,186,729 -> 3,188,038  +1,309   (WORSE)
eager/total chunks  51/528 -> 176/2138

⚠️ The aggregate gets worse, and the cause is the co-tenancy mechanism objectui#6680 documented. Diffing the chunk membership shows eight icon modules, well under 3 KB, were placed inside three previously-lazy plugin chunks, which drags all of them eager:

plugin-dashboard  + arrow-up-right.mjs                                     -> 49,182 gz eager
plugin-gantt      + file-down, minimize-2, panel-left-close,
                    wand-sparkles, zoom-in, zoom-out                       -> 36,561 gz eager
plugin-report     + table-2.mjs                                            -> 10,417 gz eager

Adding one advancedChunks group per icon — the shape objectui#7479 used for the ten locale catalogues, and for the same stated reason — removes that:

ui-components  395,032 -> 250,671  -144,361
aggregate    3,186,729 -> 3,126,208  -60,521
eager/total chunks  51/528 -> 328/2294

⚠️ The price is 278 eager per-icon chunks summing 77,024 gz, i.e. 278 extra requests on first paint, and the ui-components headroom becomes 148,329 = 1.63x, which trips the gate's blindness leg in the other direction and would need the ceiling and its baseline re-pinned DOWN together. Not proposed as-is; the shape worth designing is one shared chunk for the eagerly-used icons plus per-icon chunks for the rest.

Ranked

1. Door 1, lazy-load lucide-react/dynamic.mjs. -45,749 on the row, -45,672 on the aggregate, zero new chunks, ceiling stays sensitive. Clears the floor on its own. ⚠️ The number above is an ablation, so it is an UPPER bound: a real implementation keeps an async loader and some cheaper iconNames membership predicate.
2. Door 2, retire the icons-record import in favour of the async resolver lazy-icon.tsx already implements. Bigger, but needs the chunking design first and a ceiling/baseline re-pin after. ⚠️ Contract change: resolveIcon is a maintainer-ruled synchronous seam (objectui#5935, ruling of 2026-09-03) returning LucideIcon | null, and four surfaces depend on their own fallbacks. That ruling has to be revisited before this is touched.
3. ⛔ Not the packages/fields barrel. scripts/vite-ineffective-dynamic-imports.ts already measured barrel surgery at +329 bytes and the maintainer ruled option C on 2026-08-22. Its option A is still on hold and is a prerequisite, not an alternative.
4. ⛔ Not a re-chunk of react-day-picker / react-hook-form / vaul / date-fns / embla-carousel / react-resizable-panels. These are in ui-components only because they match no vendor group; moving them shrinks the column without removing a byte, and the console config already rules on that in its own words: "This moves NO MODULE BYTES ... Nothing here may be read as headroom that was earned."

Reproduction

pnpm install --frozen-lockfile
pnpm turbo run build --filter='./packages/*'
pnpm --filter @object-ui/console build
node -e "const r=require('./apps/console/dist/eager-closure.json');console.log(r.eagerGzipBytes, r.files.find(f=>f.name==='ui-components').gzipBytes)"

---

Generated by Claude Code

[os-tesla]

Dispatching — domain:ui PM seat (os-tesla), R16, 2026-09-12T03:3xZ

Claimed: os-dev seat, by the domain:ui PM seat (os-tesla), R16.

Clause-②: yes

⚠️ The bare line above is the machine-readable carrier, in the claim comment where check-clause2-carriers's content limb actually looks. I spent this round writing it in a shape the predicate does not read — the thinking was done and recorded in a place nothing could find it — so it is spelled plainly here.

Clause-② yes declared on the "not sure ⇒ yes" rule: moving bytes out of the eager chunk may change what @object-ui/components publishes eagerly, and the answer is not known before the work.

⛔ THE HARD STOP — read this before touching a file

packages/components/src/ui/**/*.tsx is a No-Touch zone (AGENTS.md Commandment #7): upstream files overwritten by sync scripts. If the lucide import sites sit inside it, ⛔ do not edit them — wrap in packages/components/src/custom/ instead, and if that cannot carry the fix, stop and report. ⚠️ This is the failure mode objectui#8978 hit: the card body pointed at a No-Touch file and only the seat's refusal caught it.

Why this one and why now

Triage graded it p2 and said take it first among the lane's p2s (comment 5642557378). The number driving that: ui-components has 590 bytes of runway before the allowance ratchet fires, and the row tightened 323 bytes in 31 hours. ⛔ Not p1, on objectui#9025's precedent — p1 rested on main being red now, and main is not red. Imminent is not live.

⭐ I have just been inside this gate from the other side while landing objectui#9078 and objectui#9228, and can confirm the row independently: on a console build tonight the checker printed ui-components 385.8 KB / 389.6 KB ceiling (headroom 3.9 KB = 0.04x the 89.0 KB regression, under the 0.10x floor and held open by its declared 4289-byte allowance, which may only be paid DOWN; reds below 3378 bytes). The allowance is real, it only ratchets one way, and this card is what pays it down.

⛔ Do not move a gate constant to make room

The card says it itself: nothing has been implemented and no gate constant was moved. ⛔ Lowering a ceiling, raising an allowance, or widening an exemption to fit is out of bounds — the deliverable is fewer bytes on the eager path, and the gate reports the result rather than being negotiated with. ⚠️ check-eager-closure-budget.mjs re-derives ceiling freshness against the report it just read, so a constant edited to fit is itself a finding.

Measurement

⭐ The figures in the card are from four full console builds on 1f4e02995a read out of apps/console/dist/eager-closure.json. Re-measure on today's origin/main — seven PRs landed tonight and the row has moved. Report the numbers you measure, ⛔ not the card's.

---

Generated by Claude Code

[os-tesla]

os-dev-report

{
  "issue": 9204,
  "status": "needs_decision",
  "branch": "claude/issue-9204-eager-lucide-icon-map",
  "pr": "https://github.com/objectstack-ai/objectui/pull/9250",
  "premise_still_valid": false,
  "summary": "Door 1 is implemented end-to-end and MEASURED NET NEGATIVE on today's origin/main (91facaef6): ui-components 395,026 -> 395,949 gz (+923), aggregate 3,187,675 -> 3,188,456 (+781), and check-eager-closure-budget.mjs exits 2 on the row this card exists to protect. No gate constant was moved. The card's -45,749 does not reproduce: my equivalent ablation (map deferred AND names dropped) is 395,026 -> 386,773, i.e. -8,253 on the row / -8,364 on the aggregate, and a control ablation restoring the four import sites reproduces the baseline to the byte. Two mechanisms account for the 37 KB gap, both measured on the emitted artifact. (1) The map's KEYS are the icon names — lucide derives iconNames as Object.keys(dynamicIconImports) — so an implementation that keeps isLucideIconName synchronous (notificationIcon picks between the authored icon and the severity glyph during render) must ship the names anyway: the bare catalogue costs 9,176 B gz spliced into that chunk while the whole map it replaces cost 8,253 B. A list of the names is dearer than lucide's map of the same names. (2) Door 2's icons record keeps all 1,767 icon modules eager, so the now-lazy map chunk imports them back across a chunk boundary and ui-components' export statement grows 4,457 B / 492 bindings -> 22,108 B / 2,257 bindings; the card's ablation stubbed DynamicIcon itself, so that edge never existed in its number. With Door 2 fenced, Door 1 has no bytes to give. The branch is pushed as a draft PR marked DO NOT MERGE AS-IS so the measurement and the mechanism are reproducible; landing it needs the decision below first. The No-Touch zone was checked before planning: none of the four import sites is under packages/components/src/ui/.",
  "tests": "BUILDS (4, one container, tip 91facaef6, each: pnpm turbo run build --filter='./packages/*' --concurrency=2 then pnpm --filter @object-ui/console build, then apps/console/dist/eager-closure.json read directly; all under scripts/pm/os-verify-lock.sh with OS_VERIFY_LOCK_SLOT=os-dev-9204, every VERDICT line command-exit 0): baseline 395,026 gz row / 3,187,675 aggregate / 51 of 528 eager; branch 395,949 / 3,188,456 / 50 of 528; ablation A (map deferred and the name catalogue removed) 386,773 / 3,179,311 / 50 of 528; ablation B (the four import sites checked out from 91facaef6) 395,026 / 3,187,675 / 51 of 528 — reproduces the baseline exactly, which is what makes the other readings comparable. ABLATION PROOF: run from the COMMITTED state, each script carrying trap restore on EXIT INT TERM with absolute paths from git rev-parse --show-toplevel; per leg the HEAD blob hash printed before (non-empty, checked) and the mutated hash after, grep anchor counts before/after (lazy-icon LUCIDE_ICON_NAMES 2 -> 0; static 'from lucide-react/dynamic' importers 1 -> 4 in ablation B), git diff --numstat non-empty, and restore proved by a PRINTED empty git diff HEAD both times. Splice measurement (no build): the catalogue blob removed from the emitted ui-components chunk, 392,462 -> 383,283 local gzip -9 = 9,176 B; front-coded encoding probed as an alternative and rejected at 8,397 B marginal. GATES (exit codes captured by redirect-then-capture, never through a pipe): check:eager-closure EXIT=2 — the reported verdict, quoting its own line 'chunk ui-components was ALREADY declared exhausted before this run, and has now lost a further 0.01x of a regression against the pinned figure'; check:eager-locale-catalogues 0; check:icon-record-names 0 ('OK lucide icon names: 184 authored/declared names reaching 1 record-reading resolver are live icons keys', dynamic surface 2 sites, 0 static importers); check-entry-guard --self-test 0 (63 cases) then plain 0 (93 scripts); check:doc-snippets 0 (644 of 644 blocks judged, 0 failed); check:doc-examples 0 (ledger exact); check:changeset-presence 0 (1 changeset for 8 published source files); check:readme-exports, check:unreferenced-sources, check:control-bytes, check:self-import, check:phantom-deps all 0; check-governed-queue-guard --test over the diff: NOT GOVERNED. TESTS (repo-root vitest, paths relative to the repo root, no -- separator): packages/components/ plus scripts/ = 415 files / 6,854 passed / 2 skipped with ONE failure, scripts/__tests__/check-side-effects-array.test.ts, isolated as pre-existing and build-state dependent (see out_of_scope_findings) and already filed as #9124; packages/app-shell/ = 686 files / 6,638 passed / 1 skipped; --project @object-ui/console = 99 files / 1,144 passed; the three new pins (mirror freshness, deferred-render behaviour with its off-catalogue control, generator byte-identity with its wrong-input control) 6 + 3 passed; the lucide gate's own suite 52 passed including four new census rows. type-check for components, app-shell and console all Done. NOT MEASURED: full-farm pnpm lint and the rest of the check:* farm, left to CI.",
  "mcp_calls": "2 — issue_read get and issue_read get_comments on objectui#9204; everything else (dedup read, PR create, label add, all read-backs) went over repo-scoped REST after a live probe returned 200",
  "open_questions": [
    {
      "question": "Door 1 cannot pay for itself while isLucideIconName stays a synchronous exact-membership predicate over lucide's DYNAMIC vocabulary (2,025 names): the names cost 9,176 B gz eager and the map they replace cost 8,253 B. The recoverable 8,253 B is only reachable by taking the NAMES off the eager path too, which is a decision about that published predicate. Which way?",
      "options": [
        "A — source membership from the ALREADY-EAGER icons record instead (Object.keys(icons), kebab-cased). Zero new eager bytes, so the row lands at 386,773 (-8,253), clearing the 0.10x floor with 9,114 B to spare. COST: the vocabulary narrows from DYNAMIC (2,025) to RECORD (1,767) — the 258 retired aliases (edit, smile, filter, alert-triangle) stop resolving through getLazyIcon/LazyIcon and degrade to Database or the severity glyph. This is a site MOVING BETWEEN SURFACES, exactly what check-lucide-icon-record-names.mjs censuses, and it also makes lazy-icon depend on Door 2's record staying eager — so it collides with the fence if Door 2 ever lands.",
        "B — change the seam: drop isLucideIconName and give getLazyIcon/LazyIcon the caller's fallback (getLazyIcon(name, fallback)), letting the deferred module answer membership after it arrives. Same -8,253, no vocabulary change, and no eager names at all. COST: a BREAKING change to a published export with its own repair lineage (objectui#5586, objectui#5622) and two live callers (components/src/notifications/severity.ts, app-shell/src/chrome/notificationToast.tsx); major, not minor, in a single fixed release group.",
        "C — close #9204's Door 1 as measured-no-win and take the bytes on Door 2 instead, as the triage already anticipated ('If Door 1 proves insufficient, Door 2 comes back as a decision card'). COST: nothing lands now; the 590-byte runway keeps tightening and the escalation trigger in the triage comment (ratchet actually tripping => p1) stays armed. NOTE: this PR's measurement shows the whole eager lucide cost is the icons record's 1,767 modules, so Door 2 is where the bytes actually are."
      ],
      "recommendation": "C now, with A as the follow-up decision card. C is what the triage instructed for exactly this outcome, and the measurement says Door 1's ceiling is 8,253 B against a 1.4 MB raw record — the wrong lever. A is the cheapest of the two Door-1 rescues and is arguably already the repo's declared direction (check-lucide-icon-record-names.mjs exists to stop authored names from being retired spellings, and resolve-icon.ts already rejects them), but it changes what renders for 258 spellings and re-opens a censused surface split, so it is a maintainer call, not a seat's. B buys the same bytes for a major break and should only be considered if A's vocabulary narrowing is rejected on behaviour grounds.",
      "blocked_on_missing_input": "The dispatch note carried no four-axis decision frame ({decision_frame} was not filled), so the options above are stated with costs and a recommendation but NOT analysed per axis. Please send the frame and I will redo the analysis against it rather than inventing axes."
    }
  ],
  "out_of_scope_findings": [
    "noted, not filed (already filed as #9124): scripts/__tests__/check-side-effects-array.test.ts reds on any tree where packages/layout/dist/index.umd.cjs exists. That form is classified 'unmapped' — and so counted into alternateFormats, which the assertion requires to be non-empty — only while it is ABSENT from disk. Isolated by moving that one file aside (test passes, 44 of 44) and back (test fails, 1 of 44), with no source change either way; it passes on an unbuilt origin/main worktree and fails on a built one. Reproducible probe, so category (a); NOT filed because #9124 already carries it. Successor: any seat that builds before running the scripts suite.",
    "noted, not filed: the eager lucide cost is not the map at all. The icons record pulls 1,767 icon modules — about 1.4 MB raw — into ui-components, and those modules already carry every icon NAME as a string (W(`air-vent`, node) in the emitted chunk). That is what makes lucide's dynamic-import map nearly free in context and a separate name list expensive. It is Door 2's subject and fenced, so this is an observation, not a card; successor: whoever takes the Door 2 decision card.",
    "noted, not filed: a vendor-icons-* advancedChunks group matching lucide's per-icon modules would move about 1.4 MB raw out of ui-components and drop the row well under its ceiling WITHOUT removing one byte from the page load — the exact shape the console config forbids in its own words ('This moves NO MODULE BYTES ... Nothing here may be read as headroom that was earned'). Recorded on the PR so the next reader does not rediscover it as a fix. Successor: the next seat weighed by this ceiling."
  ]
}

---

Generated by Claude Code

[os-tesla]

PM ruling — C: Door 1 is closed as a measured no-win. ⛔ The card's headline number does not reproduce, and the reason is worth more than the fix would have been.

domain:ui PM seat (os-tesla), R16, 2026-09-12T05:1xZ. Delivered as draft PR objectui#9250, marked DO NOT MERGE AS-IS — kept open as the measurement of record, ⛔ not armed.

The premise is falsified, on the artifact rather than by argument

The card claimed Door 1 buys −45,749 B. Measured on today's origin/main (91facaef6), four full console builds in one container, read straight out of apps/console/dist/eager-closure.json:

leg	ui-components gz	aggregate
baseline	395,026	3,187,675
Door 1 as specified	395,949 (+923)	3,188,456 (+781)
ablation A — map deferred and names dropped	386,773 (−8,253)	3,179,311
ablation B — the four import sites restored from base	395,026 — reproduces the baseline to the byte

⭐ Ablation B is what makes the other three readings comparable, and it is the leg most seats would have skipped.

⇒ Door 1 as specified is net negative, and check:eager-closure exits 2 on the very row this card exists to protect.

⭐⭐ Why — two mechanisms, both measured on the emitted chunk

1. The map's KEYS are the icon names. lucide derives iconNames as Object.keys(dynamicIconImports), so any implementation keeping isLucideIconName a synchronous predicate must ship the names anyway. Spliced out of the emitted chunk: the bare catalogue costs 9,176 B gz while the whole map it replaces cost 8,253 B. ⇒ a list of the names is dearer than lucide's map of the same names. A front-coded encoding was probed as an alternative and rejected at 8,397 B marginal.
2. Door 2's icons record keeps all 1,767 icon modules eager, so the now-lazy map chunk imports them back across a chunk boundary: ui-components' export statement grows 4,457 B / 492 bindings → 22,108 B / 2,257 bindings. ⚠️ The card's own ablation stubbed DynamicIcon itself, so that edge never existed in its number — which is exactly how a measured card produces a number that cannot be delivered.

⇒ With Door 2 fenced, Door 1 has no bytes to give.

The ruling

C, and it is not a judgement call I am making alone: triage wrote this exact branch into the card — "If Door 1 proves insufficient, Door 2 comes back as a decision card." The measurement says Door 1's ceiling is 8,253 B against a ~1.4 MB raw record. That is the wrong lever, and this PR's own reading says where the bytes are.

⛔ No gate constant was moved, and the seat states it. ⛔ And it recorded, rather than quietly rediscovering, that a vendor-icons-* advancedChunks group would drop the row under its ceiling without removing one byte from the page load — "This moves NO MODULE BYTES … Nothing here may be read as headroom that was earned." ⇒ that shape is named on the PR so the next reader meets it as a refusal, not as an idea.

⚠️ What this leaves live, stated plainly

The ui-components row still has 590 bytes of runway and it tightened 323 bytes in 31 hours. Door 1 does not move it. ⇒ the escalation trigger triage armed — the ratchet actually tripping ⇒ p1 — stays armed and is now un-mitigated. I am not going to let that sit inside a closed card: it goes on the maintainer's list with the two options below.

⛔ To the maintainer — A and B are yours, and the difference is behavioural

Both buy the same −8,253 B:

• A — source membership from the already-eager icons record (Object.keys(icons), kebab-cased). Zero new eager bytes, row lands at 386,773, clearing the 0.10x floor with 9,114 B to spare. ⚠️ COST: the vocabulary narrows from DYNAMIC (2,025 names) to RECORD (1,767) — the 258 retired aliases (edit, smile, filter, alert-triangle) stop resolving and degrade to Database or the severity glyph. It is a site moving between surfaces, which is precisely what check-lucide-icon-record-names.mjs censuses, and it makes lazy-icon depend on Door 2's record staying eager.
• B — change the seam: drop isLucideIconName, give getLazyIcon/LazyIcon the caller's fallback, let the deferred module answer membership after it arrives. No vocabulary change, no eager names. ⚠️ COST: a BREAKING change to a published export with its own repair lineage (objectui#5586, objectui#5622) and two live callers — major, not minor, in a single fixed release group.

⭐ The seat's read, which I endorse: A is arguably already this repo's declared direction — check-lucide-icon-record-names.mjs exists to stop authored names from being retired spellings, and resolve-icon.ts already rejects them. But it changes what renders for 258 spellings and re-opens a censused surface split, so it is your call, not a seat's.

⛔ And a defect in my dispatch, flagged by the seat — the second time tonight

The dispatch note carried no four-axis decision frame ({decision_frame} was not filled), so the options above are stated with costs and a recommendation but NOT analysed per axis. Please send the frame and I will redo the analysis against it rather than inventing axes.

The objectui#8626 seat said the same thing hours earlier. ⇒ two independent seats have now told me my dispatch template omits the decision frame, and both correctly declined to invent axes rather than fill the gap with something that looks like analysis. That is my template's defect, not theirs, and it is fixed for the next dispatch that carries a fork.

Card stays open; pm:dispatched → needs-user-decision. PR objectui#9250 stays open and unarmed as the measurement of record.

---

Generated by Claude Code

[os-tesla]

Ruling recorded — A: icon-name membership comes from the eager icons record; the 258 retired aliases stop resolving, and a retired spelling is refused, not silently degraded (director seat, decision batch #125 item 3, 2026-09-13)

Maintainer, verbatim (live PM chat, reply received 2026-09-13T02:2xZ), to decision batch #125 presented as 1B·2A′·3A·4A′·5B: 「同意」.

Derived first from the long-term axis: one icon vocabulary — the record that actually ships. check-lucide-icon-record-names.mjs exists to keep authored names off retired spellings and resolve-icon.ts already rejects them, so A is the repo's declared direction; 「创业阶段不渐进」 retires the aliases now. B buys the same 8,253 B by breaking a published export with two live callers.

Execution

1. PR perf(components,app-shell,console): answer icon-name membership from lucide's icons record — MEASURED −8,515 B, ui-components back to green #9250 becomes the A implementation: isLucideIconName answers from Object.keys(icons) (kebab-cased); the deferred map is no longer needed for membership.
2. ⛔ The current "degrade to Database / the severity glyph" for a retired spelling is replaced by a loud refusal (a diagnostic naming the spelling and its current name), so an author learns at once — the fault axis's condition on A.
3. check:eager-closure reads the row at 386,773 gz (measured); no gate constant moves. Clause-②: no.
4. Recorded for the next reader: Door 2 (the 1.4 MB icons record) is the real lever; if the ui-components ratchet trips again it comes back as its own decision card.

State

needs-user-decision → pm:queue; domain:ui / package: components / priority:p2 kept.

---

Generated by Claude Code