Collapse duplicate search rows for one symbol page

[malfet]

Follow-up to #252. That fixed the ranking half of pytorch/pytorch#4954; this is the "finds duplicates" half, which it left untouched.

On the docs preview from pytorch/pytorch#195042, searching linear returned eight rows all rendering as Linear pointing at eight different pages, and nn.linear returned four rows that were the same page reached three different ways.

Cause

performSearch runs a title pass, an object pass and a fulltext pass, each emitting its own row for a page. The built-in dedup keys on [docname, title, anchor, descr, filename] — which differs per pass, so nothing collapses:

 1. "Adam"              #adam              -> generated/torch.optim.Adam   (title pass)
 9. "torch.optim.Adam"  #torch.optim.Adam  -> generated/torch.optim.Adam   (object pass)
12. "Adam"              (no anchor)        -> generated/torch.optim.Adam   (fulltext pass)

Separately, rows are labelled with whichever heading matched — for autodoc symbol pages that's the bare leaf name, making sibling pages indistinguishable.

Fix

searchtools.js applies Scorer.score to every result before it dedupes, and hands score() the live result array. Rewriting the fields the dedup keys on therefore makes Sphinx's own dedup collapse the rows — no patching of Search.query:

var generated = /(?:^|\/)generated\/(.+)$/.exec(String(docname));
if (generated) { result[1] = generated[1]; result[2] = ""; result[3] = null; }

Relabelling to the page's own symbol fixes the display at the same time. Only generated/<dotted.path> pages are canonicalised — prose pages are left alone, since distinct sections of one page are useful deep links rather than duplicates.

Before / after

linear, before:  Linear, Linear, Linear, Linear, linear, Linear, Linear, Linear
linear, after:   torch.nn.Linear
                 torch.ao.nn.qat.Linear
                 torch.ao.nn.quantized.Linear
                 torch.nn.modules.linear.Linear
                 torch.ao.nn.qat.dynamic.Linear

Same-page duplicates in the top 10: nn.linear 4 → 0, conv2d 2 → 0, adam 1 → 0. All eight symbol queries keep their target at rank 1; cuda semantics, autograd mechanics and broadcasting semantics keep their existing top result.

Known limits

• Ordering dependency. score() must run ahead of the dedup in Search.query. That holds in Sphinx 7.2.6, which docs builds pin. If a future Sphinx reorders it, rows duplicate again exactly as they do today — it degrades to current behaviour rather than breaking. Called out in a comment in the source.
• Does not merge genuinely separate pages. torch.optim.Adam and torch.optim.adam.Adam_class are two pages documenting one class. No search change can merge those; that's an autodoc/autosummary duplication on the PyTorch side and deserves its own issue.

Alternative rejected

Switching to a document-oriented engine cannot produce same-page duplicates, since it indexes one entry per page. MiniSearch over a page-level corpus does return zero duplicates — but ranks the wanted page 3rd for adam and linear where this returns it 1st. That trades #252's ranking win for the dedup win instead of getting both.

---

Authored with the assistance of an AI coding agent.

[netlify]

✅ Deploy Preview for pytorchsphinxtheme ready!

Name	Link
🔨 Latest commit	d0f1679
🔍 Latest deploy log	https://app.netlify.com/projects/pytorchsphinxtheme/deploys/6a9182c3755e580008943eee
😎 Deploy Preview	https://deploy-preview-255--pytorchsphinxtheme.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.

---

To edit notification comments on pull requests, go to your Netlify project configuration.