v0.11.0

A correctness release from a full audit of the indexing, retrieval, CLI and
documentation surfaces. Most of it is defence: a refused deletion plan can no longer
be retried into a deletion, a truncated crawl can no longer authorize one, an
unreadable robots.txt no longer fails open, and clean can no longer delete the
project. Alongside that, chunk metadata that had been going stale forever — tags,
ordinals, route files, hook output — now reaches storage.

Three changes need action. The inert embedding.* config keys are now rejected
rather than ignored, the svelte peer floor moved to ^5.20.0, and pathPrefix
matches whole path segments. Each is marked Breaking: below.

Upgrade cost

The first indexing run after upgrading re-writes more than usual. Nothing is lost, and
the extra cost is one-off:

• Every page summary re-upserts once. The page fingerprint is now computed over
  canonical JSON, so every existing page hash differs and each page summary is written
  again.
• Unchanged chunks get a metadata-only refresh — no re-embedding. The text is not
  re-sent, so these cost a write but not an embedding.
• content-files chunks whose text contained HTML entities re-embed once. Raw
  .svelte extraction now decodes entities properly, so that text genuinely changed.

Changed

• Breaking: the inert embedding.* options are rejected instead of ignored.
  embedding.model, embedding.dimensions and embedding.taskType were validated,
  merged and typed as required, and read by nothing — embeddings are generated by
  Upstash from the text SearchSocket sends, so the model is a property of the index
  and is chosen when that index is created. They join embedding.batchSize and
  embedding.images.enable in the removed-key list, so a config that still sets one
  now fails to load with a message naming the replacement. Delete the keys; to run a
  different model, create a new Upstash index with it and point upstash.url /
  upstash.token at that index.

• A refused large-deletion plan now writes nothing at all. Previously a run stopped by
  indexing.maxDeletionRatio still upserted its additions, which enlarged the
  inventory the ratio is measured against — an identical retry then saw the same stale
  records as a smaller fraction and deleted them without --accept-large-deletion
  being supplied. The whole mutation is now refused, so a refused plan stays refused
  however often it is repeated. The separate empty-source guard is unchanged: it stops
  deletions but still writes the run's additions.

• An unreadable robots.txt blocks index writes in crawl mode, the only mode that
  fetches it over the network. If the fetch times out, returns a server error, or exceeds
  the 512 KiB size cap, the run cannot tell which pages the site means to exclude, so it
  writes nothing rather than risk publishing an excluded page. A 404 or 410 still means
  "no rules here" and indexing proceeds. static-output/build read robots.txt from
  the build directory and treat an unreadable file as absent.

• robots.txt matching follows RFC 9309 for * wildcards, the $ end-anchor, group
  merging across consecutive User-agent lines, and longest-rule-wins specificity. Not
  certified-complete: percent-encoded paths are compared as written, not canonicalised.

• llms.txt / llms-full.txt are only republished by a run holding the complete
  page list. A truncated crawl, a failed extraction, or a refused deletion plan
  leaves the existing files alone. Sharp edge: if the index holds custom records and a
  run is not passed customRecords, those pages are missing from the run's page list
  and the exports are not refreshed — the warning now says so, and names
  customRecords as the fix. Passing customRecords: [] asserts there are none.

• Breaking: pathPrefix is now a segment prefix in search. /docs matches /docs and
  /docs/api and no longer matches /docsmith. This is a behaviour change for any
  prefix that was relying on raw string matching. Note the scope: this covers the
  search and section-retrieval paths. Administrative page listing still matches on a
  raw string prefix, so listPages with /docs continues to return /docsmith.

• results: null from a search endpoint is an error, not an empty result set. The
  client and the createSearch store validate the response shape; on a 2xx response a
  null body, a parseable error page, or results: null raises Invalid search response. A rejected response is never written to the result cache.

• Title and anchor ranking keeps non-ASCII letters. Matching is Unicode-aware and
  NFC-normalized, so Japanese, Cyrillic, Greek and accented-Latin titles get the
  lexical boosts. Accents are preserved rather than folded, so a query for cafe no
  longer matches a title of café.

• Chunk metadata is budgeted in UTF-8 bytes against Upstash's 48 KB per-record
  limit, measured on the exact JSON that will be sent rather than in characters. The
  trimmable text field is truncated to fit with a warning; if the required fields alone
  do not fit, the run fails naming the page URL and the offending field.

• Raw .svelte content-files mode honours noindex/weight/ignore signals
  statically, and withholds a page when a recognised signal is present but not
  statically decidable — an expression, inside an {#if}, or a spread the scanner can tie
  to a protective attribute. Reading an undecidable protective signal as false would
  publish a page someone asked to hold back. Known gaps, pinned by tests:
  {@html}/{@render} contents are not extracted, +layout.svelte signals do not
  propagate to the pages beneath, and a conditional data-search-ignore drops its subtree
  without withholding the page.

• Breaking: the svelte peer range is now ^5.20.0. The generated templates use
  $props.id(), which landed in Svelte 5.20.0; the previous ^5.0.0 allowed versions
  that cannot compile them.

Fixed

• The relevance playground returned 500 for every error. It read error.statusCode,
  but SearchSocketError exposes status, so intended 4xx responses were reported as
  server errors.
• Documentation corrections throughout: browser examples import searchsocketScrollToText
  from searchsocket/scroll rather than the server-only searchsocket/sveltekit
  barrel; the quickstart installs search-trigger as well as search-dialog; the
  page-retrieval route is /api/search/pages/<path>, nested under api.path; hook
  examples use the real ExtractedPage and IndexStats shapes; the removed
  embedding.* keys no longer appear in a config example; scoreGapThreshold is
  described as adjacent-gap trimming, which is what the code does; scores are no longer
  described as a 0–1 confidence scale; the Upstash Vector feature claim says hybrid
  retrieval with DBSF fusion rather than reranking and input enrichment; SECURITY.md
  lists the supported versions and the anonymous mcp.handle.access: "public" mode.