docs(changelog): add the 0.12.0 entry

[jonassaa]

The changelog entry for 0.12.0, landing on main before the release exists — the ordering docs/dev/releasing.md makes step 1, because the tag is cut from main's tip at publish time and an entry merged afterwards is not in the release. That is how v0.1.2 shipped to every channel with no entry at all.

Why 0.12.0 and not 0.11.1

The policy keys the bump on the entry's own status field, and the honest value here is feature. The release is mostly performance work, which is patch territory — but #478 adds user-visible capability that did not exist before:

• a "Search all of history" control on the file-history screen,
• an on-screen notice naming how far the walk looked ("Searched the newest 50,000 commits"), with the number carried on the wire so it cannot drift from the policy applied,
• a cancel that reaches an in-process revwalk, offered from the status bar and from a new "Stop searching history" row in the command palette.

"Any new user-visible capability" is a minor pre-1.0, and the skill's red-flag list names "it's a small feature, patch is fine" explicitly. So: minor.

What the entry covers

Section	Source
New features	#478 — the cap, the notice, the escape hatch, the cancel and its routing table
Performance	#478 (135.6 s → 18.3 s on a kernel file, plus the move to the shared read lock) and #479 (the walk order prepared once, commit search included)
Improvements	#475 + #481 (the benchmark harness and the generated README block) and #477 (the site figures rendered at 2x from the real components)
Known limitations	two new, three carried forward

Two of the Known limitations are new, and both exist because this release publishes measured figures on the front page for the first time. The committed benchmark record predates the log-walk cache and was deliberately not re-measured — two sessions shared one bench directory and control rows moved 25–30%, which is two programs sharing a machine rather than a result — so the published numbers understate the app on exactly the rows the cache improves. And the ~15 s floor on opening a 1.5-million-commit repository survives both fixes, because a sorted libgit2 revwalk pre-walks the whole graph before yielding anything and the alternatives are measurably dead ends. Naming both beats letting a reader find the gap.

Pre-flight, run on this tree

Gate	Result
pnpm test	4182 passed, 407 files
pnpm tsc --noEmit	clean
cargo test	full suite, exit 0
pnpm build (site)	6 pages, exit 0
e2e	green on 50d43f5 in CI — a real run, 4 shards ~2 min each, not the no-op shim

The entry was also verified through the changelog page's own backtick transform rather than by eye: every code span pairs, none double-backticked, and the built dist/changelog/index.html carries all nine <code> spans.

This PR is site-only, so its own required checks pass as no-op shims — the e2e evidence above is the run against 50d43f5, which is the tree this release tags apart from this one file.

🤖 Generated with Claude Code