[oddharsh]

The problem

Every performance number in this repo is absolute: a measurement against a constant somebody typed. perf-budget.mjs's own baseline history is the receipt for what that costs.

baseline	outcome
86 KiB	pre-consolidation
129.23 KiB	sat 58% stale for months. The advisory fired on every run and CI printed "hard checks green" over it
204.24 KiB	set 2026-08-04 after a direct measurement. Already over its 255.30 KiB alert by 2026-08-08 (258.34 KiB at 295ee97)

Four days from a fresh baseline to a firing advisory. A constant isn't wrong because someone chose it badly; it's wrong because it describes a moment and the bundle doesn't hold still. And "258.34 against 255.30" says nothing about which PR spent the 54 KiB.

The change

The differential half, modelled on how astral-sh/ruff runs its memory and ecosystem jobs: build the merge base, build HEAD, measure both, post the delta, gate on nothing.

• scripts/perf-snapshot.mjs records the built tree's wire sizes (11 client assets, 46 pages, the /a/ immutable tier, worker bundle gzip, per-module esbuild attribution, dictionary-delta coverage) and diffs two recordings into markdown. One wrangler deploy --dry-run self-builds, so a snapshot costs ~12s.
• .github/workflows/perf-diff.yml runs it per PR and updates one marker-keyed comment, matching the pattern already in dependabot-site-review.yml.

It is deliberately not part of validate. validate is the one required check on main, and a perf number that blocks a merge teaches people to widen the threshold rather than to read it.

What the report looks like

A tooling-only PR (this one):

Client assets

No change. 11 files, 56.82 KiB Brotli total.

Pages

No change. 46 files, 465.36 KiB Brotli total.

Worker bundle

No change. 258.34 KiB gzip.

That silence is the feature. A report that always has content stops being read.

A one-line nav.js edit:

Client assets

Total Brotli: 56.82 KiB → 56.87 KiB (+0.05 KiB, +0.1%) across 11 files, 1 changed.

file	base	head	Δ
nav.js	13.41 KiB	13.46 KiB	+0.05 KiB (+0.4%)

Pages

Total Brotli: 465.36 KiB → 465.33 KiB (-0.03 KiB, -0.0%) across 46 files, 38 changed.

No single file moved by 0.13 KiB or more.

_{Plus 38 files under the 0.13 KiB floor, net -0.03 KiB — usually content-hash churn in /a/ references, counted in the total above.}

Three design decisions worth reviewing

Both sides are measured with HEAD's copy of the script, stashed in $RUNNER_TEMP before the first git switch. The base doesn't have the script on the PR that adds it, and measuring each side with its own copy would report a change to the measurement as a change in wire size.

The noise floor is asymmetric: pages get 128 bytes, client assets get none. A page carries /a/<name>.<hash8>.<ext> references, so touching one shared asset flips its hash and moves every page's compressed size (measured: 38 of 46 pages, max 41 bytes, net -0.03 KiB). An asset's bytes are its own content, so nothing but editing it can move them and every byte is signal. One floor everywhere would have hidden the 50-byte nav.js change that caused the churn. Sub-floor movers collapse into a counted aggregate line and stay in every total.

Nothing sampled is recorded. wrangler check startup read 9.6, 7.6, 6.4 and now 16.4 ms across identical bytes, a 2.6x spread. Diffing two draws from that manufactures findings in a comment whose credibility rests on every row being real. That check is reframed from "regression tripwire" to diagnostic; the tripwire is the deterministic bundle gzip and module attribution next to it. Same instinct as CodSpeed counting simulated instructions and ty's memory job pinning TY_MAX_PARALLELISM=1 — remove the noise source rather than widen a threshold around it.

Deliberately not done

The worker advisory is left firing rather than re-baselined. Re-baselining an unattributed 54 KiB is indistinguishable from silencing the check, and the diff is the thing that can now name which PR spent it. The observation is recorded in the baseline history instead.

Verification

Run locally against origin/main:

• npm test — 179 pass
• npm run perf-budget — hard checks green (1 advisory, the pre-existing bundle one)
• npm run typecheck, npm run check-wrangler, npm run pages:check — clean
• All five workflow YAMLs parse
• End-to-end: recorded origin/main, recorded this branch, diffed — silent, as it should be for a tooling-only change
• Synthetic 50-byte nav.js edit produces the table above, then reverted

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	aadhar-sh	00d7579	Aug 08 2026, 08:19 PM

[github-actions]

Wire-size diff vs merge base

2f781a6 → 0d0d39a

Brotli q11, the quality the edge serves. Every number here is deterministic:
an unchanged file produces no row. Advisory — this check fails on nothing.

Client assets

No change. 11 files, 57.99 KiB Brotli total.

Pages

No change. 46 files, 465.79 KiB Brotli total.

Worker bundle

No change. 261.74 KiB gzip.

Dictionary deltas (pd/)

No change. 134 deltas, 373.59 KiB total.