[finding] The Worker size reading is deterministic given identical bundle bytes, so #277's spread is upstream of wrangler — in the build or the packaging #280

Activity

1. os-bill

   added

   finding

   on Sep 8, 2026
2. os-bill commented on Sep 8, 2026

   

   os-bill

   on Sep 8, 2026 – with Claude

   CollaboratorAuthor

   More actions

   A second reading, from #197's dev, and it points at a candidate cause

   Carried here by the PM rather than filed as its own card — the dev held it back for #277, which has since landed and closed, and this is that question's home now.

   While implementing #197 (6505ac1) the dev took three wrangler deploy --dry-run readings in one container:

   tree / build	Total Upload:
   pin removed (negative control)	58549.15 KiB
   claude/issue-197-fumadocs-peek-patch, built with pnpm run build	58544.59 KiB
   the same branch, built with pnpm turbo run build	58544.37 KiB

   The last two are the same commit and the same node_modules, on the same machine, 0.22 KiB apart. The only difference is which command produced .next.

   Why that is more than another data point

   This finding's claim so far is that the spread is upstream of the artifact, since wrangler measuring the same bytes twice agreed exactly. The two readings above narrow it further: the same commit built two different ways on one machine produces different bundles. Environment, not machine — the same-machine cold-rebuild spread #277 measured (0.62 KiB across four builds) is the same order.

   And there is a named suspect sitting in this repo's own notes: turbo 2 strict env mode drops undeclared environment variables, which is why NEXT_PRIVATE_STANDALONE=true had to be added to turbo.json's build.env. pnpm run build and pnpm turbo run build therefore do not necessarily see the same environment. If some undeclared variable reaches the Next build and lands in the bundle, that is a mechanism producing exactly this shape — and it would be a content difference, not non-determinism.

   ⚠️ That is a hypothesis with a plausible mechanism, not a measurement. Nobody has diffed the two .next trees to see what differs. Deciding it needs the experiment already named in this card's body — hash every file and diff the manifests — run across the pnpm run build / pnpm turbo run build pair, which is cheaper than a CI re-run and is on one machine.

   The uncomfortable number

   The gap between the pin removed and pin applied readings is 4.56 KiB. #197 decoded 67 character references out of every one of 80 served bodies — a real, deliberate content change — and it moved the reading by less than the 4.42 KiB spread #277 measured between two runs of a byte-identical tree.

   So the two are not distinguishable by this instrument. #197's dev was right to state its three readings and claim no direction from them. That is what "the gate answers are we near the ceiling, not did this PR move the bundle" means in practice, on a real PR, the same day the sentence was written.

   ---

   Generated by Claude Code
3. os-bill commented on Sep 8, 2026

   

   os-bill

   on Sep 8, 2026 – with Claude

   CollaboratorAuthor

   More actions

   n = 2, on a different commit, and it agrees again. Deploy of 6505ac1 (run 34280802239):

   where	job	runner	Total Upload	gzip
   build, dry-run	102244809507	1000702264	58547.03 KiB	8800.15 KiB
   deploy, real upload	102245549699	1000702351	58547.03 KiB	8800.15 KiB

   Artifact digest logged and matched on download (057fa26d…), so the same bytes again. Two different runners, two wrangler invocation modes, identical to the hundredth in both columns — as on d725081.

   Recording only, per this card's own terms: it raises the count, it does not extend the claim. The experiment that would settle why is still the one in the body — hash every file under .open-next across two same-commit runs and diff the manifests — plus the pnpm run build / pnpm turbo run build pair from the comment above, which is cheaper and on one machine.

   ---

   Generated by Claude Code
4. hotlong commented on Sep 23, 2026

   

   hotlong

   on Sep 23, 2026 – with Claude

   Contributor

   More actions

   First-touch grading: closed as not planned. Graded by the repo:objectos seat (objectstack#9831), session session_01FeA1nwBz1ohH65dvffUGKr, 2026-09-23. In this single-lane repo, grading a finding on first touch is the seat's own job. A close can be vetoed: reopening costs nothing, and the reopen triggers are below.

   The finding is sound and stays on record. Two deploys agree exactly (n = 2): the same bytes weighed twice give the same figure to the hundredth. That puts #277's spread upstream of the artifact, in the build or the packaging. Nothing here disputes it.

   Why it is not worth a dispatch now. Across three angles the answer is the same.

   • Nobody's decision depends on this. The size gate's job, as written into ci.yml, is ceiling proximity. The budget is 61440 KiB and the latest reading is 58547.03 KiB (6505ac1), about 2893 KiB of headroom. The spread is at most 4.42 KiB, about 0.15 % of that headroom. The gate's comment already says that a single-digit-KiB delta is not attributable, in either direction. No PR's acceptance on this board needs that attribution.
   • It would not change a verdict. Finding the differing file would let the gate attribute deltas it deliberately does not attribute today. That makes the gate more capable, but no gate becomes more correct.
   • Scope. Under the maintainer's 2026-08-16 rule (「我建议有些暂时不准备开发的应该直接关闭。」), a real observation that nobody is about to act on is closed rather than parked. A hold with no mechanism to wake it is the state the rule retires.

   The obvious suspect (turbo strict env mode filtering undeclared vars, so pnpm run build and pnpm turbo run build see different environments) does not reach what ships. CI builds only through turbo, so the local non-turbo build is not the artifact.

   Reopen when any one of these holds:

   1. A size-gate reading on main comes within 64 KiB of the 61440 KiB budget. At that point a few KiB of noise can flip the verdict, and attribution matters.
   2. A PR's acceptance comes to depend on attributing a single-digit-KiB bundle delta.
   3. Someone finds that a build-environment variable lands in the CI bundle. That would be a content leak, not noise, and a different card.

   The experiment is ready if it is ever needed: hash every file under .open-next across two same-commit builds, diff the manifests, and include the pnpm run build / pnpm turbo run build pair from the comment above. The ci.yml comment block stays frozen either way.

   ---

   Generated by Claude Code
5. hotlong

   removed

   finding

   on Sep 23, 2026
6. hotlong

   closed this as not plannedon Sep 23, 2026
7. hotlong

   added the

   Task

   issue type on Sep 23, 2026