Fixes #1390

When a message only changes capitalization or surrounding whitespace, its fuzzy key can already be in the candidate dictionary. This checks that dictionary first instead of comparing the text with every candidate again.

The existing search gives identical normalized strings a score of 1, so the shortcut picks the same key. The merge code is unchanged: translations, fuzzy flags, comments, previous IDs and obsolete messages still go through the same path. Actual wording changes still use get_close_matches(). No extra cache is added.

For Q exact-normalized lookups among C candidates, this avoids Q × C candidate comparisons and uses Q dictionary lookups instead, after the existing candidate preparation. String normalization still takes time, and nonexact fuzzy matching has the same complexity as before. This does not make every catalog update linear.

This leaves the current normalized-key collision behavior alone; #1330 handles that separately.

Checks:

• 419 before/after comparisons matched, covering update options, plurals, Unicode, empty normalized text, collisions, comments, flags, previous IDs, obsolete messages and template contents.
• The full suite ran on both implementations. PR: 7,835 passed; base: 7,834 passed. Both had the same two failures in test_get_timezone_name_misc on this Windows machine (GMT+00:00 versus GMT+05:30), plus 9 skips and 2 expected failures. The base run excludes the new test that specifically checks fuzzy-search call elimination.
• Ruff and git diff --check pass. The full messages tests pass on both versions.

Benchmark scripts, raw samples and reproduction steps.

Measurements use Python 3.14.5 on Windows. Each row uses one fixed pair of physical cores; pairs rotate between rows. These Python functions are single-threaded, not parallel workers. Results from different pairs are shown separately because this laptop has different core types.

Each row has five paired timing samples per version (short calls batched), plus five separate tracemalloc measurements. Times and peaks are medians. GC stays enabled. Peaks cover newly traced Python allocations during the call, not total process RAM or a copy of the installed library.

The full workflow parses existing PO bytes, updates from a prepared template, and writes the entire PO. Reading files and preparing the template are outside the measurement. The catalogs are real Django/Wagtail translations; the edits are controlled capitalization, surrounding-whitespace, or punctuation changes, not production traffic.

All output bytes matched. The unchanged and punctuation-change controls do not use the shortcut, so their timing differences are not claimed as improvements. Every measured row is included, including slower results.

Base commit: e2a1791. Source, runner and fixture hashes and every raw sample are in the linked JSON.

Fixtures (downloaded from these pinned sources; not copied into the benchmark gist):

• Django German
• Wagtail French