perf(localized_reader): O(1) name lookup via HashMap index (#191)

[SAY-5]

Closes #191.

scan_directory looked up the existing entry for a normalised full_name via entries.iter_mut().find(|e| e.name == full_name) on every file processed. With n total entries that's an O(n²) walk — fine for tiny prompt trees, painful once a workspace accumulates a few hundred .src.mdx / .mdx pairs across the zh / en / dist combinations.

Fix

Add a side-table HashMap<String, usize> mapping the normalised name to the entry's index in the Vec. The dedup branch becomes by_name.get(&full_name) (O(1)), and the new-entry branch inserts (full_name.clone(), entries.len()) before the push so the next lookup hits.

The seen HashSet is left in place to keep the diff minimal — it's addressed separately in #190 (PR #242).

Test plan

• cargo build --manifest-path sdk/Cargo.toml — clean
• cargo test --manifest-path sdk/Cargo.toml --lib repositories — 45/45 pass

[TrueNine]

Thanks for the contribution on #253. Retargeting this to dev on our side, and I am adding explicit regression coverage so the behavior stays pinned down while we take the lookup optimization.

[TrueNine]

Thanks again for the contribution on #253.

I applied the lookup optimization onto dev on our side and added explicit regression coverage to pin down the intended behavior while taking the performance improvement.

What we added:

• a regression test documenting that #253 keeps localized variants grouped into the same FlatFileEntry
• coverage that a zh source, an en source, and a compiled mdx file with the same normalized name still merge into one entry

I verified the result with:
cargo test --manifest-path sdk/Cargo.toml --lib repositories

Current result: 46 passed.

So the optimization is now covered by a behavior-preserving test and is included on our dev branch for the next release. Since the change has already been applied manually on our side, I am closing this PR to keep the branch history clean. Thanks for the careful contribution.