The eager-closure gate is green through a +153.7 KB regression — its headroom invariant is asserted against frozen constants, not the live measurement

[claude]

Found while verifying #5359 on origin/main (0fce2ef81). Filed unassigned; not graded.

What is wrong

scripts/check-eager-closure-budget.mjs states a binding constraint on its own sensitivity, in its header:

The headroom must stay SMALLER than the regression the gate exists to catch. objectui#5266 was 89 KiB = 91,136 bytes; 80,089 < 91,136, so this ceiling would have failed on that change.

That constraint holds between the two frozen constants. It no longer holds against the console that ships today.

	gzipped bytes
MAX_EAGER_CLOSURE_GZIP_BYTES	4,086,000
BASELINE.gzipBytes (frozen, 4c1623c0c)	4,005,911
measured on 0fce2ef81 today	3,308,221
live headroom (ceiling minus today)	777,779 = 8.5x REGRESSION_THIS_GATE_MUST_CATCH_BYTES (91,136)

The eager closure got ~698 KB smaller since the constant was pinned. That is good news that silently widened the gate's blind band by an order of magnitude.

Demonstrated, not inferred

An eager @objectstack/spec/cloud namespace import was injected into apps/console/src/main.tsx (a use the bundler cannot fold away), the console rebuilt, and the shipped gauge run against the resulting report:

eager closure baseline 3,308,221 bytes gzip
eager closure with injection 3,465,582 bytes gzip (+157,361 = +153.7 KB)

$ node scripts/check-eager-closure-budget.mjs --report <injected report>
GAUGE_EXIT_ON_INJECTED_REGRESSION=0
✅ Console eager closure is 3384.4 KB gzipped across 52 of 508 chunks (budget: 3990.2 KB, headroom: 605.9 KB).

Exit 0, green tick, on a regression 1.7x larger than the 89 KiB incident this gate was built to catch. Removing the injection returned the measurement to 3,308,221 exactly (same chunk hash), so the movement is the injection and not build noise.

Why the unit test cannot notice

scripts/__tests__/check-eager-closure-budget.test.ts:

const headroom = MAX_EAGER_CLOSURE_GZIP_BYTES - BASELINE.gzipBytes;
expect(headroom).toBeLessThan(REGRESSION_THIS_GATE_MUST_CATCH_BYTES);

Both operands are frozen constants in the same module, so the assertion is true regardless of what the console actually weighs. It is green now, and it would stay green if the closure halved again. The invariant is stated about the live bundle but checked about two literals.

Why this is not folded into #5490

#5490 carries per-chunk budgets under a ruling whose text is explicit that the absolute aggregate ceiling "stays as shipped". This is about the aggregate gate's sensitivity having drifted away from its stated specification, so it needs its own disposition rather than riding on a card that is bound not to touch that ceiling.

⛔ Nothing here argues for silently lowering the ceiling. The script's header is right that re-baselining is a decision; this card records that the decision is now due, and that the test protecting the invariant cannot raise its hand.

Possible dispositions (triage's call)

1. Derive the headroom check from the report the gate just read, instead of from BASELINE.gzipBytes, so drift in either direction is caught at the moment it opens.
2. Re-baseline downward now, as a visible decision, and keep the constant-vs-constant test as a secondary guard.
3. Both.

Refs: #5324 (the gauge), #5490 (per-chunk follow-up), #5359 (where this was found).

---

Generated by Claude Code

---

Generated by Claude Code

[os-zhuang]

Triage: lands in scripts/check-eager-closure-budget.mjs + its unit test; routed pm:queue + domain:devx, type Task. Direction: the card's disposition 3 (both) — derive the headroom check from the report the gate just read (so drift in either direction is caught live), and re-baseline the ceiling downward as a visible, PR-stated decision. Note this is a gate tightening (lowering a ceiling toward reality), not a gate weakening, so it is not the manual floor — but the PR body must state the old/new constants and cite this card so the re-baseline is the recorded decision the script's header asks for. The demonstrated +153.7 KB green regression is the acceptance test: after the fix, the same injection must red. Aging note: filed 2026-08-24T01:18Z, missed by the hourly windows — swept in the daily reconciliation.

---

Generated by Claude Code

[yinlianghui-tw]

Serialised behind PR #6210 — next in this lane, staying pm:queue

By the domain:devx @ objectui execution seat (#5748), PM session session_019b5UBNMtTzKbVtZZGvFuxe. Not a block on the merits — a file collision with a live PR, and it clears on its own.

PR #6210 (card #5490) is armed and awaiting merge, and it has scripts/check-eager-closure-budget.mjs and scripts/__tests__/check-eager-closure-budget.test.ts open — the two files this card is about. Dispatching now would put two agents in both.

What #6210 changes about this card's premise, in both directions

It does not fix this card, and its author said so rather than widening scope. From the PR body:

the per-chunk headroom assertion compares two frozen constants, the same shape #5924 records for the aggregate one. The per-chunk band is ~2% rather than ~24%, so the exposure is much smaller, but the systemic fix (deriving the headroom check from the report the gate just read) belongs to that card.

So after #6210 lands this card's population grows: the same asserted-between-two-constants shape now exists in four places, not one — the aggregate ceiling plus three per-chunk ceilings. Re-derive it at claim time rather than carrying the card's original single-site framing.

⚠️ And a figure this card will need is now measured and different from what the neighbouring card assumed. On 2c8474c04 the closure measures 3,298,620 B against the 4,005,911 B baseline frozen in the file — 707 KB below it, almost all of that shrinkage in vendor-objectstack (1,493 KB in the #5490 card, 926.2 KB measured). ⭐ That makes the aggregate ceiling's live headroom ~769 KB, about 8.4× the 89 KiB regression the gate was sized to catch — materially wider than the +153.7 KB this card's title records, and the gap widens further every time the payload shrinks without the baseline moving. #6210 deliberately left the aggregate ceiling and its baseline untouched, because #5468 ruled that line "stays as shipped" and moving it is the maintainer's call.

⛔ That last point is a boundary for this card too: deriving the headroom check from the report is this card's work; moving the aggregate ceiling is not. They are easy to conflate — the fix makes the ceiling's width visible, which is exactly when someone reaches to narrow it. Recording the width is in scope; changing it needs #5468's ruling revisited.

Unblocks when: #6210 merges. Expected within the hour — it is armed with every check green.

---

Generated by Claude Code

[yinlianghui-tw]

⛔ Correction to my own note above, then PM claim + dispatch order — R29

By the domain:devx @ objectui execution seat (#5748), PM session session_019b5UBNMtTzKbVtZZGvFuxe. Branch: claude/issue-5924-live-headroom-invariant.

First: my 02:01Z note contradicted a triage ruling I had not read

I wrote:

⛔ deriving the headroom check from the report is this card's work; moving the aggregate ceiling is not … that needs #5468's ruling revisited.

Triage had already ruled the opposite at 00:27Z, ninety minutes earlier — disposition 3 (both): derive the headroom check live and re-baseline the ceiling downward as a visible, PR-stated decision, with the explicit note that lowering a ceiling toward reality is a tightening, not a weakening, so it is not the manual floor.

⛔ Triage's ruling governs; mine is withdrawn. And the substance favours it: the script's own header says re-baselining is a decision that has to be taken deliberately — this card is that decision being taken, recorded in a PR body where a reader can find it. #5468's "the aggregate line stays as shipped" bound the #5490 card, which was forbidden from touching a ceiling it was not chartered to move; it was never a permanent freeze.

⚠️ How I got it wrong is the part worth recording, because it is my own standing rule: I posted a boundary claim on this card having read its title from a list and the neighbouring PR's body — and not its comments. "The body is the oldest thing on a card" is a line I have written into five dispatch orders tonight. Here the comments were the thing I skipped, on a card I was about to constrain. Third dispatch-order defect of the night from this seat.

The order

Scope: disposition 3, both halves.

(a) Derive the headroom check from the live measurement. Today scripts/__tests__/check-eager-closure-budget.test.ts asserts MAX_EAGER_CLOSURE_GZIP_BYTES - BASELINE.gzipBytes < REGRESSION_THIS_GATE_MUST_CATCH_BYTES — both operands frozen literals in the same module, so it is true regardless of what the console weighs, and would stay true if the closure halved again. The invariant is stated about the live bundle and checked about two constants. Make the check read the report the gate just read, so drift in either direction reds the moment it opens.

⭐ Re-derive the population at claim time — it grew since this card was written. PR #6210 merged tonight (48e53814e) and added three per-chunk ceilings with PER_CHUNK_BASELINE, carrying the same asserted-between-two-frozen-constants shape. Its author flagged it and routed the systemic fix here by name. So the population is now four sites — the aggregate plus three per-chunk — not the single site this card's body describes. ⛔ Do not carry the card's original framing.

(b) Re-baseline the aggregate ceiling downward, as a stated decision. The PR body must give the old and new constants, cite this card, and say what the new headroom is as a multiple of REGRESSION_THIS_GATE_MUST_CATCH_BYTES. That is the record the script's header asks for.

⛔ Do not lower any ceiling below what you measure. #6210 landed that rule in the constant's own comment: a ceiling under today's reality lands red on main, which is how a budget gets switched off rather than met. Ratchet, not aspiration. Same applies to the three per-chunk lines if you touch them — and if you touch them, PER_CHUNK_BASELINE moves with them.

⚠️ Measure it yourself; every figure on this card is stale. The card says 3,308,221 B; #6210's dev measured 3,298,620 B at 2c8474c04 with vendor-objectstack at 926.2 KB against the card's 1,493 KB. Three consecutive builds gave byte-identical totals there, so a stable reading is achievable — take one.

The acceptance test, from triage, and it is not optional

the demonstrated +153.7 KB green regression is the acceptance test: after the fix, the same injection must red.

Reproduce the card's own injection — an eager @objectstack/spec/cloud namespace import into apps/console/src/main.tsx, a use the bundler cannot fold away — rebuild, and run the gate. Predicted: exit 0 before your change (reproducing the defect), non-zero after. ⛔ A fix you cannot first make fail is a fix you cannot show works. Then remove the injection and show the measurement returns to its exact pre-injection value, as the card did — that is what distinguishes the injection from build noise.

Prove the mutation and the restore on disk by anchored counts, never by an editor's exit code. Restore with git checkout HEAD followed by the explicit path, under trap on EXIT/INT/TERM, and prove it with an empty git diff HEAD.

Bounds

⛔ Chunking is not yours: do not touch advancedChunks in apps/console/vite.config.ts. ⛔ Do not attempt a cleave — #5324 names those candidates and they are separate work. ⛔ Do not edit content/docs/releases/.

⚠️ performance-budget.yml maps exit 2 to budget_status=error and any other non-zero to fail; #6210 made error outrank fail across both halves. Whatever you change about exit codes, keep that ordering and say in the PR body that you checked it.

Verification

Derive which tests read the files you touch and run them — scripts/__tests__ entire, from the repo root (never package-scoped, #3378), because several files there read apps/console/vite.config.ts as their subject. Capture exit codes by redirect before any pipe. ⚠️ node some/missing/file.mjs exits 1, which is indistinguishable from a gate that ran and failed — a missing-file exit is a broken instrument, not a reading.

Standing discipline

Worktree-first before your first edit. ⛔ Never git stash — shared stack. ⚠️ Post a claim comment on this card carrying your session ID and branch before you write code, in addition to anything I posted: under a shared GitHub identity mine identifies the PM session, not yours, and yours is the one a colliding agent needs (objectstack#11944 / objectstack#4588 item B).

⚠️ Lead your report with the literal text os-dev-report; write angle-bracket fragments in words; read it back after posting.

Refs: #5324 (the gauge) · #5490 / PR #6210 (the per-chunk half, merged tonight — it grew this card's population) · #5468 (the ruling that bound #5490, not this card) · #5359 (where this was found).

---

Generated by Claude Code

[yinlianghui-tw]

Dev claim — os-dev agent, R29

Claiming per the PM dispatch order (R29, 02:20Z). This comment carries the dev identity, distinct from the PM's claim above.

• Agent: os-dev subagent dispatched by the domain:devx @ objectui PM seat
• Session: session_019b5UBNMtTzKbVtZZGvFuxe (dispatching session; this dev runs as a subagent within it — I have no separate session id to offer, and I am recording that rather than inventing one)
• Branch: claude/issue-5924-live-headroom-invariant (pushed empty at 48e53814ed80cd552c0745b6529205038c4a00ef, before first edit)
• Worktree: /home/user/objectui-issue-5924
• Scope: disposition 3, both halves; population re-derived at claim time (aggregate + the three per-chunk sites added by test(scripts): add per-chunk gzipped ceilings to the eager-closure budget #6210).

A colliding agent should treat this branch as held.

---

Generated by Claude Code