fix(search): relevance ranking, per-version de-dupe, product-scoped version filter

[aidenhongnet]

Problem

The full search page re-sorted every Algolia hit alphabetically by product before paginating, so the top-ranked result for a query could land pages deep while low-scoring hits from alphabetically-early products always filled page one. The Ctrl+K preview had the same problem from a different cause: DocSearch buckets hits by hierarchy.lvl0, which the crawler sets to "Product > Version", so the preview grouped by product and dropped results past a per-group cap. Both surfaces hid Algolia's ranking.

Two further problems on the same surfaces (second round of commits): the same article appeared once per product version — the index legitimately stores one record set per (product, version) copy of a page, and both surfaces rendered them all — and the Versions filter rendered even with no product selected; on /search it listed the union of all 27 products' versions, which is meaningless without a product.

What changed

Search results page — now server-side, relevance-ordered.

• Renders Algolia's relevance order directly. The alphabetical-by-product sort is gone; product filtering stays available via the product/version facet dropdowns.
• Replaced the fetch-1000-hits-and-paginate-client-side design with Algolia server-side pagination (hitsPerPage = page size, setPage(n)). Per-search payload dropped from ~4.7 MB to ~170 KB.
• A stale/overshooting ?page=N recovers to the first page instead of rendering an empty, control-less dead end.
• Banner reflects the reachable cap: "the top 1,000 of N — refine" when more match than Algolia will page to, exact count otherwise.

Preview (Ctrl+K modal) — now relevance-ordered.

• ungroup() collapses DocSearch's two grouping passes (hierarchy.lvl0 and hierarchy.lvl1) so hits render in Algolia's order; the real title is fed back through _snippetResult (Algolia-escaped, highlighting preserved) and the product is shown per hit.
• This is a client-side shim around the crawler putting the product in lvl0. The root-cause fix is to set lvl0 to a constant in the Algolia crawler config, after which the shim can be deleted — noted in-code.

Version de-dupe — one row per article, newest version wins (both surfaces).

• Each fetched page collapses duplicate articles to the version-newest copy, in its original ranking slot. Identity = the hit's URL minus #fragment minus its version path segment, with trailing-slash variants normalized, standalone /docs/kb/... pages rewritten to their versioned-copy shape, and version-pinned KB source basenames (kbSource overrides, e.g. accessanalyzer-2601) normalized to the product id.
• "Newest" is config order in products.js (versions[0] is the unique isLatest for all 27 products) — digit parsing is provably wrong here (passwordreset's latest 3.3 predates legacy hidden 3.23 numerically). Versions missing from config rank oldest.
• Selecting any specific version disables de-dupe — per-version rows return, restricted to the selected versions. The SearchPage gates on the facet carried by the response's own request, so in-flight filter changes can never mis-gate.
• New shared src/theme/searchUtils.js holds the de-dupe and version-list logic; it replaces two diverged local copies of getVersionsForProducts.

Version filter — only meaningful with a product selected (both surfaces).

• The Versions dropdown renders only while at least one specific product is selected, listing only the selected products' versions, newest first.
• The same rule gates the query facet and the versions URL param, so stale bookmarked URLs or leftover sessionStorage can never filter invisibly while the panel is hidden.
• The modal passes an extended attributesToRetrieve (DocSearch replaces its defaults wholesale, so they are re-listed alongside product_name/product_version — pinned with a re-verify note for future DocSearch bumps).

Adjacent fixes on the same surfaces.

• Re-enabled typo tolerance (was hard-disabled on both surfaces; any misspelling returned zero results).
• Added distinct: true to the page helper (deduped near-duplicate hits; matches the modal).
• Removed advancedSyntax (a query like Get-ADUser -Filter silently excluded docs containing "Filter").
• Moved the Algolia result listener into a useEffect with cleanup (was re-registering every render with stale closures); added a real debounce; fixed a <meta property="robots"> → name="robots" typo.
• Deleted dead code the refactor surfaced: the config transformItems (never serialized), an unused clsx import, an unused CSS module, and disjunctiveFacets.

Verification (de-dupe/filter commits)

All nine planned dev-server flows pass: modal and /search de-dupe, product-scoped version lists, version selection restoring per-version rows, stale versions= URL scrubbing with de-dupe intact, modal→page filter handoff, and identitymanager (current wins), passwordreset (3.3 beats legacy 3.23), and KB standalone/versioned merge spot-checks. A standalone logic suite exercises the de-dupe key against the real products.js. An adversarial multi-agent review of the diff confirmed one finding (the pinned-KB landing slug divergence), which is fixed in the second commit.

Known follow-ups (not in this PR)

• KB pages under /docs/kb/{general,recoveryad,tags}/ have empty product_name in the index and badge as "Documentation" — a crawler/index gap.
• The ungroup shim depends on DocSearch internals; a future @docsearch/react major upgrade could silently revert the preview to product order (documented in-code with a re-check note). The crawler lvl0 fix removes this dependency entirely.
• Cross-product KB duplicates (the same KB article copied into different products) remain separate results — distinct product docs; the product filter covers it. Likewise the legacy shared docs/kb/recoveryad standalone set, which backs two products and has no unambiguous merge target.
• nbHits/banner counts still count all versions (fixing that needs global knowledge of the result set; de-dupe is per fetched page by design).

🤖 Generated with Claude Code