📝 docs: summarize perf gains and refresh the benchmark tables

[gaborbernat]

A single changelog entry for the 1.5.0 cycle's performance work, summarizing the C-core optimizations landed since 1.4.0 (#648 through #680) rather than leaving them without a user-facing note. The numbers come from a same-machine A/B of the release branch against the 1.4.0 tag on a fixed workload: an id-selective XPath lookup about 260x faster, overlapping text search over 100x, Unicode normalization about 6x, and a spread of validation, extraction, and minification operations 2x to 5x, alongside the serialize scan's ~24% gain. The entry groups the changes by technique so a reader sees the shape of the work, not a flat list of issue numbers.

Publishing those gains surfaced a defect in the bench tooling. report.py names each emitted feed after its operation, but the performance guide names its tables for the section they sit in (querying-2.json, building-3.json), and nothing mapped the two. So --table-json docs/development/bench could never refresh the committed feeds; it dropped operation-named files beside them, and every refresh was a hand copy. That is how the guide's tables went stale and quietly lost competitor columns, while report.py's own docstring claimed a straight copy was all it took. bench.docs_feeds now holds that mapping and writes the committed feeds, mirroring bench.migration, which already did the job for the per-library feeds, and a test pins the map against the guide's own bench-table directives so the two cannot drift apart again.

Every feed is regenerated from a full profile-guided run through that generator, so the committed files are reproducible output rather than a hand copy, and the refresh procedure is written into the guide instead of being folklore. The read-path tables pick up the wider competitor field the suites gained, and the prose across the performance and migration guides is rewritten so each ratio, competitor list, and comparative claim tracks a measured number. Claims that had flipped are corrected: turbohtml now leads libxml2 on XML parsing and libxslt on XSLT, lightningcss on every stylesheet it accepts, and parsel no longer sits out the text collectors. The competitor timeout also moves from 900s to 5400s, so a library that reparses the page on every mutation iteration, or walks multi-megabyte input, gets measured instead of dropped for being slow.

[codspeed-hq]

Merging this PR will not alter performance

✅ 106 untouched benchmarks
⏩ 20 skipped benchmarks¹

---

_{Comparing gaborbernat:docs/perf-changelog-1.5 (c956f53) with main (e8af0a7)}

Footnotes

1. 20 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩