v0.2.0

The data model moves
from "ship every snapshot in the tarball" to
"cache on first use, fetch from the hosted Worker, fall back to a
bundled subset." The npm version stops tracking spec data and starts
meaning code again.

Added

• ECMA-402 reaches edition parity with ECMA-262. 402 publishes
  each annual edition as an esYYYY branch (not a tag); the catalog
  now exposes the full es2016 – es2025 range plus main, where it
  previously had only es2025-candidate + main. spec.diff,
  spec.history, and edition-pinned clause.get now work across the
  402 annual line.
• ECMA-402 proposals are indexed. proposal.list / proposal.get
  now cover the ecma402/ proposal set (~32 proposals that were
  entirely missing), and proposal.list gains a spec filter
  ({ spec: "402" }). Every proposal row carries a spec tag.

Removed

• Dropped the transient es2025-candidate 402 pin. It predated the
  final es2025 branch, which now supersedes it. clause.get,
  spec.diff, and the other edition-aware tools no longer accept
  edition: "es2025-candidate" — use es2025.

Changed

• Snapshots are sourced through a cache → hosted Worker → bundled
  fallback chain (loadSnapshot). The stdio server caches each
  snapshot under ~/.cache/tc39-mcp/ on first network fetch and serves
  it from disk thereafter, revalidating against the Worker only after a
  ~4-hour freshness window (conditional If-None-Match).
• latest on ECMA-402 now resolves to es2025 (the newest annual
  edition), matching 262's "latest = newest stable" semantics. Use
  main / draft / next for the working draft.
• The npm tarball shrinks ~70%. It now bundles only the offline
  fallback — spec-262-es2025, spec-262-main, spec-402-es2025,
  spec-402-main, and the proposals + test262 indexes — instead of
  every parsed edition. All other editions are fetched from the Worker
  on demand.
• cheerio and @tc39/ecma262-biblio move to devDependencies.
  They're only used by the parser at build time (npm run parse); the
  running server reads pre-parsed JSON and never imports them. A
  consumer install no longer pulls cheerio's ~20-package HTML-parsing
  tree (~2.4 MB). The published tarball also drops dist/docs and
  dist/refresh (build/CI-only code that runs from src/ via tsx,
  never at runtime).
• Refresh decouples from the npm version. refresh.yml updates R2
  every ~4 hours (live freshness for networked users) and re-bakes the
  npm bundle at most monthly, ending the ~2000-PATCH-bumps-per-year
  flood. A new annual edition still publishes immediately via the
  normal code-release path.
• @tc39/ecma262-biblio is pinned to an exact version (per its
  README's "pin a precise version" guidance), so the parse metadata
  layer is reproducible alongside the SHA-pinned spec HTML.
• The 262 parser gained an HTML-discovery fallback: since the
  pinned biblio is a snapshot of main, it can lag the HTML being
  parsed (a newer main, or an older edition carrying clauses since
  dropped). Any <emu-clause> the biblio doesn't list is now still
  captured, with metadata synthesized from the element — so a stale or
  mismatched biblio can no longer silently drop a clause.

Fixed

• Conditional revalidation returns 304 from the hosted Worker. The
  stdio loader sends If-None-Match when re-checking a live snapshot
  past its ~4-hour freshness window; the /r2/ proxy previously ignored
  it and re-sent the full body. It now returns a bodyless 304 on an
  etag match, so a revalidation costs a header round-trip instead of a
  tens-of-MB re-download.

Notes

• stdio now makes HTTPS requests to source non-cached snapshots
  (by default the hosted Worker; override with TC39_MCP_BASE_URL to
  self-host or air-gap). Tool arguments and clause ids never leave the
  process — see docs/privacy.md.
• v0.1.x installs are unaffected; this is a backward-compatible minor
  at the tool/protocol level.