[daggerok]

What

• scripts/update-data.ts: every file the updater writes (index.json, meta.json, holdings and history pages, update-state.json) goes through feedJson (compact JSON plus a trailing newline); holdings and history pages carry rows as string arrays in headers order (pageRows) instead of objects keyed by the header
• src/main.tsx: fetchPage reads array rows and still reads object rows
• scripts/update-data.test.ts: the zero-diff pipeline test also checks that every written file is one line and that holdings and history pages carry array rows in header order
• README feed layout describes the page shape
• api/ishares: all 5,523 files rewritten into the new format (second commit)

Why

ETF-H3: the feed was 788 MiB of the 1 GB GitHub Pages site limit and grows weekly; at the limit the Pages deploy fails. Indentation was 28% of it, object rows repeating every header text another 40%

Sizes

	before	after
api/ (bytes of the 5,523 JSON files)	787.6 MiB	254.5 MiB (-67.7%)
api/ on disk (du -sk)	817,520 KB	271,772 KB
built dist/ (du)	799 MiB	266 MiB (271,952 KB)

Data rewrite

A one-off Bun script (kept out of the repo, in the ignored node_modules/.work/) imports feedJson and pageRows from the updater and rewrites every JSON file of api/ishares through them (tmp file + rename). Before writing it checks that each file is exactly JSON.stringify(_, null, 2) + "\n" output (so the compact form is a lossless round trip), that page headers are unique and that no row has a key outside headers (0 problems over 5,039 pages and 2,314,923 rows). A second script compared the old tree (git archive HEAD api) with the new one: same file list, non-page files deep-equal, page rows read by the new reader equal the old rows read by the old reader for all 5,523 files

Consumers (all checked)

• iShares app (src/main.tsx fetchPage): reads both shapes (this PR)
• ETFs hub (fetchPage): reads both shapes since feat(sheets): read holdings and history rows published as arrays ETFs#86, merged and deployed to Pages before this PR
• investments shell: only maps /ETFs/<Brand>/api/** to files, does not parse them
• ui-std browser tests read the normalized rows of the apps, not the files
• the other 28 brands are untouched and keep object rows

How it was checked

• bun install -E, bun test (36 pass), bun build --target=bun scripts/update-data.ts --outfile=/dev/null, bun build --target=bun src/main.tsx --outfile=/dev/null, bun run build, bun run build-github-pages, git diff --check
• .claude/tools/etfs/tc/check.sh: the same 53 pre-existing lines as on main, no new error
• ui-std/brand-uitest.ts on the built dist (holdings and history tabs load the array pages): 44 passed, 0 failed; console errors: none
• live run TICKERS=AAXJ,IVV,AGG REQUEST_SLEEP=0 bun scripts/update-data.ts on the rewritten feed: 3 unchanged, zero git diff (the writer produces the rewritten bytes exactly, index.json included)
• hub in local mode over this feed (investments/scripts/serve.ts, /ETFs/?api=local, headless Chrome): AAXJ, AGG and IVV holdings and history cells equal the files, 29 passed, 0 failed; console errors: none