refactor!: remove record-adapter-sqljs and the FTS4 branch it required

[cuibonobo]

Summary

Removes @haverstack/record-adapter-sqljs, the FTS4 code that existed only to serve it, and the browser-support claims it backed. Net −2288 lines.

sql.js was chosen for runtimes without full Node support — principally Cloudflare Workers — and never served that case:

• initialize()/open() called initSqlJs() with no options. Workers forbid compiling WASM from bytes at runtime; the only route is injecting a pre-compiled WebAssembly.Module through Emscripten's instantiateWasm, and the adapter exposed no options pass-through at all. The same gap left a bundled browser app unable to set locateFile.
• Even with the WASM injection fixed, persist() serialized the whole database after every write. In a per-request Worker isolate that has no single-writer guarantee, against docs/spec/adapters.md § Concurrency & storage ownership.
• Its tests ran under environment: 'node', so the package only ever demonstrated sql.js working in the one runtime its own module comment said it was not for.

With the adapter gone, FTS4 has no consumer. Removed: FTS4_SCHEMA_SQL, fts.ts, the FtsStrategy seam (which existed solely because FTS4 and FTS5 differ on unindexing), and the FTS4→FTS5 migration in record-adapter-sqlite plus its test block. SharedSqlRecordLogic now reaches fts5Strategy directly instead of taking fts and sanitizeSearch as constructor deps, and buildWhereClause(query) calls sanitizeFts5Query rather than threading a sanitizer through.

Browser support is removed from the README package tables and repo tree, CONTRIBUTING's tree, and the adapters spec rather than caveated — the claim was what created the obligation reported in the issue.

Closes #144. Follow-ups filed: #160 (blob-adapter-s3), #161 (record-adapter-do-sqlite).

Spec

Observable behavior changes — an adapter is gone, and record-adapter-sqlite no longer opens FTS4 files. docs/spec/adapters.md is updated in the same change:

• Adapter backends — record-adapter-sqljs row dropped from the table
• Adapter backends (sqlite-shared paragraph) — rewritten for a single consumer, plus a new paragraph recording why SqlExecutor is synchronous and what that rules out (D1, libsql over HTTP). The File compatibility note now states the file carries an FTS5 index, so a reader needs FTS5 and not merely SQLite.
• Adapter capabilities — sql.js removed from the contentFieldQuery local-adapter list, the per-adapter notes, and the maxAttachmentBytes: null list
• Concurrency & storage ownership — the record-adapter-sqljs bullet (persist-callback delegation to the embedding host) is deleted

docs/spec/wire-format.md § Bounding query cost is also touched, as fallout from merging main — see below.

docs/spec/attachments.md is untouched: nothing here changes the blob contract or garbage collection.

Verification

All five, on the post-merge tree:

pnpm run format:check   # clean
pnpm run lint           # clean, 8 packages
pnpm test               # 1053 passed across 7 packages
pnpm run build          # clean
pnpm run typecheck      # clean

Checked by hand: grep -ri "sqljs|sql\.js|fts4" returns nothing outside pnpm-lock.yaml, and the surviving browser mentions are all unrelated (wire-format's HTTP-client note, id.ts/did.ts runtime-agnostic claims, the bookmark commons doc). pnpm install regenerated the lockfile.

Notes for reviewers

Merged main (#162) into this branch, which conflicted in three places:

• record-adapter-sqljs/src/index.ts and sqlite-shared/src/fts.ts — modify/delete. Main added maxContentBytes to the sqljs adapter's capabilities and rewrote the FTS4 sanitizer's cost note; both files stay deleted and neither change has anywhere to land.
• sqlite-shared/src/fts5.ts — both sides rewrote the module comment's query-timeout paragraph. Took main's version, which points at the new Bounding query cost spec section, minus its reference to the FTS4 sanitizer as a peer.

That new spec section also described both engines blocking in-process as the reason a timeout can't be set from inside the call. With sql.js gone only node:sqlite is left to describe, so this PR edits one sentence of wire-format.md that main just added — flagging it because it's outside the diff you'd expect from a package removal.

Slightly beyond the literal FTS4 deletion: buildWhereClause's sanitizeSearch parameter and its SanitizeSearch type are gone too. Once FTS4 went, it had exactly one caller passing one constant and no test stub — dead injection rather than a seam. Happy to restore it if you'd rather keep the shape.

SqlExecutor stays, even though sqlite-shared now has one consumer. Its synchronous shape is what would make #161 cheap, and the spec now says so explicitly instead of leaving it implicit.

Not in this PR: @haverstack/record-adapter-sqljs@0.6.0 is still live on npm. This drops the publish: script but can't touch what's published — deprecating the package with a pointer to record-adapter-sqlite is a manual step.