Render the public curriculum shell without delayed mobile LCP

[sarthakagrawal927]

Fleet PSI Swarm reproduced a severe public-homepage regression: mobile-mid p75 LCP 6.40s, mobile-fast p75 4.27s, and mobile-slow p75 15.80s. The LCP is curriculum-card copy and 84-92% of the delay is render delay.

Acceptance:

• preserve the public curriculum journey and authentication behavior
• render the public shell in the initial response rather than waiting on client work
• mobile-mid p75 LCP <= 2.5s across a five-run swarm
• CLS <= 0.1

Tracked from the Fleet Console performance remediation pass.

[sarthakagrawal927]

Local implementation is complete on fix/issue-42-initial-shell at commit b9dec60 (not pushed or deployed per this repository's local agent instructions).

The fix makes the generated curriculum summary the above-the-fold initial-response shell only on /, and keeps it in place until the lazy destination route has committed. Existing / → /today, guest, and sign-in behavior remains unchanged.

Validation:

• live five-run baseline: mobile-mid p75 LCP 8.369 s, CLS 0.010
• local production build, five mobile-mid runs: p75 LCP 1.965 s, CLS 0.021, performance score 98
• pnpm test: 55 files / 375 tests passed
• pnpm lint: passed
• pnpm typecheck: passed
• pnpm docs:validate: passed
• pnpm exec vite build: passed
• full pnpm build: curriculum generation passed, then correctly stopped because VITE_GOOGLE_CLIENT_ID is not available in the agent environment; no secret or production config was accessed

Remaining gate: push/review/deploy the commit, then repeat the five-run swarm against production before closing this issue.

[sarthakagrawal927]

Verified against production and closing.

#66 shipped the initial-response shell; #74 removed a shift it introduced. Both are live on main@5213c79, deployed via the Deploy to Cloudflare Pages workflow (run 31889430472, no migrations).

Acceptance, measured on production

Five-run psi-swarm against https://learn.significanthobbies.com/:

preset	p75 LCP	p75 CLS	Perf
mobile-mid	1.22 s	0.010	100
desktop	656 ms	0.002	100

• preserve the public curriculum journey and authentication behavior — unchanged; / → /dashboard handoff verified in-browser at 1350×940 and 390×844
• render the public shell in the initial response rather than waiting on client work — the LCP element is shell copy (<p> on mobile, the <h1> on desktop), not React output
• mobile-mid p75 LCP ≤ 2.5 s — 1.22 s, against the 8.37 s baseline recorded above
• CLS ≤ 0.1 — 0.010 mobile, 0.002 desktop

What #74 fixed

The first production verification of this issue passed on mobile but showed desktop p75 CLS 0.117, which failed the ≤ 0.1 bar. index.css is loaded asynchronously, so the shell painted under UA defaults and reflowed the instant Tailwind's preflight applied. Lighthouse attributed the entire shift to one node:

0.1125  div#lcp-shell > div > div#public-entry-shell > section

The dominant contributor was box-sizing — the shell's max-width:64rem + padding block laid out content-box, then rewrapped to border-box. The font stack, line-height, and the paragraph/heading/list margin resets accounted for the rest. #74 pins those values in the head under #lcp-shell <tag> selectors, which outrank preflight regardless of load order, and adds a regression test so removing them fails CI.

Desktop performance score also moved 96 → 100.

Remaining, not blocking this issue

The top opportunity on both presets is now ~110 KiB of unused JavaScript on the entry chunk — a bundle concern, unrelated to the shell, and worth its own issue if it becomes a priority.