Build perf phase 2: move meta/annotations to public/, add hashed manifest, delete per-proof shims

[rjwalters]

Phase 2 of the durable fix for #20984 (architect proposal: #20984 (comment)). Blocked by phase 1 (#20992) — do not start until phase 1 has landed and proven out the static-data + fetch pattern.

Goal

Collapse the remaining ~7300 data modules (2435 index.ts shims + 2435 meta.json + 2435 annotations.json) to a few hundred app modules by moving JSON to public/ and deleting the import.meta.glob machinery. Returns the build toward its ~35s baseline and makes it O(1) in gallery size.

Implementation (per architect proposal §1–3)

1. Emit step: copy each proof's meta.json + annotations.json → public/data/proofs/<slug>/ (alongside source.lean from phase 1). Compute sha8 hashes.
2. Hashed manifest: emit src/data/proofs/data-manifest.json = { slug: { meta: <sha8>, ann: <sha8>, src: <sha8> } }. This is the ONLY per-proof data still imported into the graph (one small module). Bundled with content hashing, so any proof change busts the importing chunk.
3. Rewrite getProofAsync to fetch meta.json/annotations.json/source.lean by slug with ?v=<sha8> query params from the manifest (see proposal loader sketch). Keep the public signature.
4. Delete the ~2435 per-proof src/data/proofs/<slug>/index.ts shims and the import.meta.glob('./**/index.ts') + proofModules map. Also delete dead code getProof/getAllProofs (0 external consumers per architect grep).
5. listings.json stays an in-graph static import (1.8MB, single module, needed eagerly by HomePage/ErdosPage/ProfileModal — not part of the O(N) blowup).

Acceptance

• pnpm build vite build stage back near ~35s baseline; report wall-clock.
• Module count in the build is a few hundred, not ~12K (vite/rollup output or --debug confirms).
• Proof pages render meta + annotations + source — browser smoke test.
• Cache-busting verified: changing one proof's meta.json changes its ?v= and the manifest chunk hash.
• getProof/getAllProofs/proofModules removed; no remaining import.meta.glob for proofs.

Follow-up

Phase 3 (#FOLLOWUP): apply the identical pattern to src/data/research/index.ts (1982 research problems, import('./problems/${slug}.json') — same disease).

[rjwalters]

Phase chain: blocked by phase 1 #20992; phase 3 = #20994.