docs(scripts): record what a re-baseline absorbs, and that BASELINE.commit cannot be checked from main (objectui#7848) - #9208
Conversation
…ommit cannot be checked from `main` (objectui#7848) Prose only. No constant moves in either direction and no verdict logic changes: the diff is comments in `scripts/check-eager-closure-budget.mjs`. The file re-pins `BASELINE` by an absolute fresh reading of a named tree, never by a delta. That is the right construction, and it has a consequence the file states nowhere: a fresh reading is silent about its own composition, so every byte that arrived since the previous baseline and was not removed in between is inside the new constant with no name on it, while the record above that constant names only the cause that motivated the move. Ruling item (2) of 2026-09-11 asks for that to be written down. Three things go in. 1. The mechanism, plus the one re-baseline here whose record can be audited. objectui#7479's control/after table closes its own accounting: the aggregate fell 410,553 while the catalogue chunk alone fell 415,781, and the remaining 5,228 is not absorbed non-locale content — it is `i18n-runtime` (+5,329), a new eager member of the closure that the same change created, leaving 101 bytes moving the other way. The eager chunk counts agree independently (`d8b4739d4` 50 of 518, `755d34a5f` 51 of 528: of ten new chunks exactly one is eager). What makes that auditable is the three unmoved control rows, not the total. 2. What was carried in anyway. objectui#7848 measured 3,263,896 on `52cac3886` against 3,222,314 on `3d257c85a` — 41,582 unattributed bytes — and `52cac3886` is an ancestor of `755d34a5f` (`ahead_by=627, behind_by=0`), so a large part of it is inside the constant. A floor under that figure is refused and the refusal is recorded: the one large removal in the span is the catalogue change, and this file's own record has the catalogue growing across the drift window (446,076 on `177afeba1` to 454,602 on `bbe285ee7`), so drift that is unattributed cannot also be asserted disjoint from the chunk that left. And the tree that was measured is not the tree that landed: after `755d34a5f` that branch took 22 more commits before the squash merge `77b2a18a16`, three of which edit non-test source inside `framework`'s own group. 3. Provenance. `BASELINE.commit` is a branch tip; `git fetch origin 755d34a` answers "couldn't find remote ref" and `git merge-base --is-ancestor` cannot resolve the object. That is the convention working, not a defect — but the consequence is that the provenance of this constant cannot be checked from a `main` checkout with git alone, and the file now says so and names the compare API and the reachable squash merge. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01FhBNJcLRZLe8M87VcUgpKr
✅ 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
|
ACCEPT — objectui#7848 ruling item (2) / PR #9208Reviewed against the diff at 1. Verified in the tree
2. ⛔ The round refuted the starred centrepiece of its own dispatch, and it was right toI told this round that The full retraction is on the card, at the altitude the claim was made: #7848 (comment) — it originated in the item-(1) report, I endorsed it as "a second absorption nobody had recorded", and I starred it in the dispatch. Three seats, two rows of a five-row table, nobody read to the bottom. ⭐ And the round found the rule my own instruction was missing. I wrote "re-derive every figure yourself… if one does not reproduce, that is the most valuable thing you can report." The four subtractions did reproduce, to the byte. In the round's words: "Arithmetic reproducing is not the same as an inference holding, and the instruction asks for the subtractions. What found the defect was reading the table the subtractions were taken from." ⇒ ⛔ re-deriving a figure is not re-deriving a claim. That goes into the dispatch template above the old line. 3. Why the PR is right to record less than it was told toTwo of the three things I ordered recorded did not survive re-derivation, and the PR says so instead of writing them:
⭐ And what it recorded beyond the order is the best part: the accounting is published with its rows, its residual (−101), and the three unmoved control rows plus both chunk counts that make it auditable — with the observation that a record publishing only its new constant is not falsifiable at all. One clause landed in 4. Routed, ⛔ not folded in — objectui#9209, and it re-opens a number I accepted
⇒ 5. GatesUnit test 138/138, the six other tests that read this source 226/226, ⭐ And the ablation is the right one: it mutates out the two commit hashes the new prose anchors and shows objectui#7528's pin go red, naming both new counts — ⛔ 6. Card stateobjectui#7848 stays open — agreed, and for the round's reason: both ruled items are executed, but that is not the card's closing condition. Its first sentence is about a live margin, and Proceeding to the merge queue. Generated by Claude Code |
Armed — auto-merge
|
| anchor | <merge>^ |
<merge> |
|---|---|---|
What a re-baseline ABSORBS |
0 | 2 |
| line count of the file | 2,448 | 2,573 (+125) |
⭐ The heading reads 2, not 1 — measured rather than assumed, because it appears in the header's index as well as at the section.
L5 — nothing moved in transit, by blob. scripts/check-eager-closure-budget.mjs on the merge commit = 895a2c64caa4243e07fc71e0731118b5fc3a7286. ⭐ Free cross-check: this is the same blob prefix the report quotes as the pre-mutation baseline of its ablation (895a2c64 → 09c9c907). The round's ablation and my read of the tree describe one object.
L6 — the prose is HELD BY A TEST, and this leg is why that matters. objectui#7528's pin requires every chunk count in this file to be anchored by a commit hash. Both anchors, two-sided:
| anchor | <merge>^ |
<merge> |
|---|---|---|
50 of 518 |
1 | 2 |
51 of 528 |
0 | 1 |
⇒ the new accounting introduces a chunk count that did not appear in this file before, and the pin the round ablated is what keeps it anchored. ⛔ A documentation PR whose sentences nothing tests is a promise; this one is not.
⛔ NOT MEASURED: the gate's own run and the seven test files. This seat has a bare, stale checkout — and the lesson of the last round is that a runnable gate here answers fluently about the wrong tree. CI's run on this head is the reading.
A leg that comes back amber will be published with its cause. ⛔ It will not be re-run into green.
Generated by Claude Code
LANDED —
|
| leg | reading |
|---|---|
| L1 file set | 1 — scripts/check-eager-closure-budget.mjs ✅ |
| L2 prose only | numstat = 125 0, and added lines that are not comment-or-blank = 0 ✅ — the predicate stated, not the claim asserted |
| L3 ⛔ NOT ONE CONSTANT MOVED | all 17 constant lines appear on both sides of the merge, byte-identical ✅ |
| L4 what moved | What a re-baseline ABSORBS 0 → 2; file 2,448 → 2,573 lines ✅ |
| L5 nothing moved in transit | blob 895a2c64caa4243e07fc71e0731118b5fc3a7286 ✅ — the same object the round's ablation used as its baseline |
| L6 the prose is held by a test | 50 of 518 1 → 2, 51 of 528 0 → 1 ✅ |
⭐ L3 is the leg this whole round existed under, and it is why the constants' line numbers moving (MAX_EAGER_CLOSURE_GZIP_BYTES 371 → 474) is not a finding: the leg compares strings, never a path:line. 125 lines landed above them; nothing landed in them.
⭐ L6 is what makes this more than a promise. objectui#7528's pin requires every chunk count in this file to be anchored by a commit hash, the new accounting introduces a count that was not in the file before, and the round ablated that pin to show it go red naming both counts. A documentation PR whose sentences nothing tests is a wish.
⛔ The correction this round carried, restated so the PR record holds it
The dispatch's starred centrepiece — +5,228 B of non-locale content absorbed in a single commit — does not exist. i18n-runtime is a new eager chunk carved out by the same change (- → 5,329); the i18n family nets −410,452 and the residual non-i18n movement is −101 bytes. That re-baseline absorbed essentially nothing.
The figure originated in the item-(1) report, I endorsed it, and I starred it in the dispatch. Full retraction on the card: #7848 (comment)
⭐ And the rule it produced, in the round's own words: "the four subtractions DID reproduce, to the byte. Arithmetic reproducing is not the same as an inference holding… What found the defect was reading the table the subtractions were taken from." ⇒ ⛔ re-deriving a figure is not re-deriving a claim.
⇒ the PR correctly records less than it was told to: the 41,582 is not written as a floor (its third leg asserts disjointness about bytes the card calls unattributed — you cannot hold both), and the provenance caveat is split rather than flattened (34a1578ef exits 1 and genuinely is not an ancestor; 755d34a5f exits 128 — the absence of a reading, not a negative one).
Card state
⛔ objectui#7848 stays open. Both ruled items are executed, but that is not the card's closing condition — its first sentence is about a live margin, and ui-components sits ~590 bytes from firing. ⛔ Closing on item (2) would be objectui#8241's shape, which is why objectui#8554 exists. The slimming route is filed as objectui#9204; the branch/PR gap this round found is objectui#9209.
Generated by Claude Code
2 participants
Part of #7848 — ruling item (2) of 2026-09-11T11:54Z. ⛔ This does not close the card: item (1) is answered, but the condition #7848 describes is still live (
ui-componentssits ~590 bytes from firing, and the slimming route is objectui#9204).⛔ Prose only. The diff is 125 added comment lines in
scripts/check-eager-closure-budget.mjsand nothing else. No constant moves in either direction — notMAX_EAGER_CLOSURE_GZIP_BYTES, notBASELINE, not anyPER_CHUNK_GZIP_CEILINGSorPER_CHUNK_BASELINEentry, notEXHAUSTED_HEADROOM_FLOOR_MULTIPLEor its granularity, notREGRESSION_THIS_GATE_MUST_CATCH_BYTES, not the'ui-components': 4_289allowance. No verdict logic changes. Not a line of executable behaviour.What goes in
1. The mechanism, and the one re-baseline here whose record can be audited.
BASELINEis re-pinned by an absolute fresh reading of a named tree, never by a delta — the right construction, with a consequence the file states nowhere: a fresh reading is silent about its own composition, so every byte that arrived since the previous baseline and was not removed in between is inside the new constant with no name on it, while the record written above that constant names only the cause that motivated the move.2. What was carried in anyway, with the chain from #7848 re-derived, and an explicit refusal to write a floor under it.
3. The provenance caveat.
BASELINE.commitis a branch tip; a reader who runsgit show BASELINE.commitfrom amaincheckout gets nothing and, until this PR, had no way to learn from the file why.⭐ The correction — one of the three figures I was dispatched to record does not mean what the dispatch says it means
The dispatch and the ACCEPT above it both carry this as "the stronger half":
Re-derived rather than copied. The arithmetic reproduces exactly and the inference does not. The answer is in the same table the two subtractions come from:
i18n-runtime, a new eager memberi18n-runtimeis the i18n provider, hooks and formatters, carved out by the same change and staying eager. It does not exist in the control column at all, it is i18n, and it accounts for 5,329 of those 5,228 — leaving 101 bytes moving the other way. So that re-baseline absorbed essentially nothing; it is the clean case, not the incriminating one.Two independent confirmations, neither of them the subtraction:
d8b4739d4weighed 50 of 518 chunks,755d34a5f51 of 528 — ten new chunks, exactly one of them eager. Nine deferred catalogues plusi18n-runtimeis the only assignment that fits.apps/console/vite.config.ts,emitEagerClosureReportsetseagerChunkCount: files.lengthandeagerGzipBytesas the sum overfiles, so any per-chunk figure the report carries is by construction an eager member.i18n-runtimeappears in the "after" column as a report reading.⇒ the PR records the accounting rather than the inference, and records what made it auditable: three unmoved control rows and both chunk counts. A re-baseline that publishes only its new constant cannot be audited at all — not because its number is wrong, but because nothing in it is falsifiable. That is the actionable half, and it is now a clause in "## Raising it".
The chain reproduces where it is checkable:
52cac3886is an ancestor of755d34a5f— compare reportsahead_by=627, behind_by=0, merge base52cac3886. ✅d8b4739d4to755d34a5fisahead_by=1. ✅ The docblock's same-container one-commit A/B claim is independently confirmed.But the third leg of the argument — "#9122 removed locale-catalogue bytes, a chunk disjoint from the drift" — is an assertion about bytes that are, by the card's own words, unattributed, and this file's own record contradicts its premise:
i18n-localesmeasured 446,076 on177afeba1(2026-09-03) and 454,602 onbbe285ee7(2026-09-10), +8,526 bytes, and that window contains52cac3886(2026-09-05). No reading of the catalogue exists on3d257c85aor on52cac3886to separate the two.⇒ the file now says a large part of the 41,582 is inside the constant, and says why a floor is refused. Writing "at least 41,582" would be exactly the unbound figure objectui#8964 exists to refuse.
⭐ Third — a gap nobody on the card has named: the measured tree is not the tree that landed
755d34a5fis the first commit of the branch, not the pull request. PR #9122's head isd7d6c6a688, 22 commits further on, squash-merged as77b2a18a16; one of those commits (bf268c3723) merged 15 more commits ofmainin. Three of the 22 edit non-test source insideframework's own group (packages/(core|react|types)), 175 added lines against 47 removed:Control with a known direction that hits in the same run — the same filter over the same 79-file compare, test files instead of source, which must be non-empty or the filter is matching nothing:
⛔ What that costs gzipped is unmeasured, not small — nobody has built
77b2a18a16, and this round deliberately spends no console build. TheBASELINEdocblock already warns that CI weighs the merge ref while this is the branch tree, "so the two differ by whatever has landed onmainsince"; the half it did not cover is that they also differ by whatever landed on the branch after the reading was taken. The PR adds the actionable rule: take the reading on the tree you are about to merge, or name the commits that arrived after it.Provenance, recorded as a consequence — ⛔ not as a fix
git fetch origin 755d34a5f1answers "couldn't find remote ref", andgit merge-base --is-ancestor 755d34a5f1 origin/mainexits 128 — the object does not resolve, which is ⛔ not the exit 1 that would mean "resolved, and not an ancestor". The previous34a1578efdoes resolve and is genuinely not an ancestor: exit 1, read against a control legd9580f4647(2026-09-06T01:40Z, five minutes newer than34a1578ef, so at least as deep) that exits 0 in the same shallow checkout — an exit 1 from a shallow clone means nothing without one.⛔ Both are squash-merged branch tips and that is the file's deliberate convention working, not an error, and nothing here proposes changing it. The consequence is what nobody had written down: the provenance of the number this board argues about cannot be checked from a⚠️ explicitly never a substitute for the reading.
maincheckout with git alone. The file now says so, names the compare API as what does resolve these shas, and names the reachable squash merge77b2a18a16as a handle on what landed —Verification
Exit codes by redirect-then-capture, ⛔ never through a pipe; verdict lines quoted from the tool, ⛔ never a bare
$?.vitest run --project unit scripts/__tests__/check-eager-closure-budget.test.tsTest Files 1 passed (1)·Tests 138 passed (138)check-eager-closure-budgetTest Files 6 passed (6)·Tests 226 passed (226)pnpm type-check:scriptstsc -p tsconfig.scripts.json, no diagnosticseslint scripts/check-eager-closure-budget.mjs --no-inline-config --format jsonnode scripts/check-lint-coverage.mjslint coverage: 46/46 packages linted, 0 with outstanding errors (0 total).node scripts/check-changeset-presence.mjsNo source or published contract of a released package changed in this range, so no changeset is owed.node scripts/check-control-bytes.mjscheck-control-bytes: OK (scanned 7400 tracked text file(s); skipped 85 binary).node scripts/check-new-cross-file-line-citations.mjsVERDICT new-cross-file-line-citations: 0 new citation(s), enforcement report-only -> exit 0node scripts/check-governed-queue-guard.mjs --test scripts/check-eager-closure-budget.mjsNOT GOVERNED — 1 path(s) checked against 5 governed surface(s); none matched.⭐ Ablation — the objectui#7528 pin really does read the two chunk counts this PR adds. Run from the committed state,
HEADcarrying the change; mutation proved on disk before the run (HEADblob895a2c64→ mutated09c9c907, both grep anchors printed verbatim, injected spelling counted 0 before and 2 after); restore isgit checkout HEAD -- pathundertrap ... EXIT INT TERMwith an absolute root, ⛔ never a baregit checkout -- path.Removing the two commit hashes from that one paragraph — and nothing else — turns the pin red and it names both counts:
Restored and proved by state, ⛔ not by an exit code:
git diff HEADempty, and the file's blob back to895a2c64— byte-identical toHEAD.Declared narrowings (⛔ these are declared deviations, not silent ones; the farm runs on CI):
Test (shard N/4): ran the seven test files that read this source rather than the whole sharded suite. Enumerated bygrep -rl check-eager-closure-budget scripts/__tests__plus the two consumers ofscripts/__tests__/helpers/attached-docs.ts— both already in the set.Lint: linted the one file the diff touches. Population read from ESLint's own--format jsonoutput over the root scope: 336 files, of which this diff touches 1. Invariance:eslint.config.jsconfigures no type-aware linting (noparserOptions.project, noprojectService), so a comment-only edit in one.mjscannot move the verdict on any untouched file. The root scope's single pre-existing error ise2e/live/global-setup.tsno-console, untouched by this diff and outsidepnpm lint's turbo scope.Build Docs: its first step is "Check for docs changes" and this diff changes no docs input — declared to CI.All figures above were re-derived in this round from the file, from
apps/console/vite.config.ts, and from the compare API. ⛔ None were copied from the dispatch comment or from the item-(1) report.🤖 Generated with Claude Code
https://claude.ai/code/session_01FhBNJcLRZLe8M87VcUgpKr
Generated by Claude Code