zit 0.13.0

zit 0.13.0 aligns its line diff and three-way merge with Fossil's, and makes the reads a forge pages through cheap on large stores: annotate, the history of one file, and the daemon's start and feed.

This release is from trunk. Stores and the wire are unchanged from 0.12.1. Two caches beside each store are rebuilt once; see Compatibility.

Diff and merge follow Fossil

• Line diff. Every line diff in zit, from diff and its line counts to annotate and merge, is a port of the line diff in Fossil's diff.c (D. Richard Hipp). It replaces the patience and Myers engines, so zit aligns repeated lines as Fossil does, and annotate credits a check-in only with the lines its diff adds.
• Three-way merge. merge walks the two edit scripts as Fossil's merge3.c does. An edit beside the other side's change now merges cleanly. Where both sides add lines at the same point, the merge now stops with a conflict, as Fossil's does, instead of taking both. Every merge decision and merged text matches fossil 3-way-merge, except where zit keeps a final newline or a byte-order mark on purpose.

annotate

• -n N is Fossil's version limit. annotate -n N walks at most N versions of the file, newest first. A line the oldest walked version already carried is printed with blank check-in and date columns, as fossil annotate prints it. Before, -n counted check-ins and credited older lines to the oldest check-in shown. A script that read those columns from every line must now allow blank ones, or pass -n none.
• Versions come from the file-history index, and the walk stops as soon as -n or the nearest cached version allows. annotate -n 50 of the most-edited file in a 644 MB store went from 38 s to 50 ms.
• A blame cache. annotate keeps the answers for a file's recently annotated versions in .zit-cache and starts from the nearest one, so annotating a tip after a new commit costs tens of milliseconds instead of seconds. An answer that depended on a missing parent, a gap, a limit or content removed by policy is never cached.

log for a path

• zit log PATH reads content only for the rows it prints and decides the rest from the file index. A 200-row page of a 2,000-version history costs tens of milliseconds.
• --offset N skips N rows, as Fossil's finfo spells it. --cursor C starts after the row a cursor names, and -F json returns the next cursor. A cursor page costs the same at any depth; an offset page still costs the rows it skips. Page by cursor where you can.

Daemon

• Start. zit daemon answers its first request in about 0.3 s on a 68,000-file store, where it took 8 to 10 s. It opens what a commit needs on the first commit, from the saved log index, and that commit costs about a second more than later ones.
• After a commit the daemon reads only the frames the commit added instead of reopening the store.
• Feed pages. A feed request may carry limit, the most check-ins to return; the reply's more says whether more follow. A page of 500 costs about 0.2 s. A request without limit is answered as before, and every feed reply now carries more. more can be true when only tag or other control artifacts follow, in which case the next page is empty. A daemon older than 0.13.0 refuses a request carrying limit.
• A session that reads the whole bootstrap feed peaks at 1.0 to 1.1 GB on a 68,000-file store, paged or not.

Compatibility

Stores, check-ins, the sync protocol and the daemon protocol version are unchanged.

Two caches in .zit-cache are rebuilt from the store, because the line diff behind them changed: the blame cache, by the first annotate of each file, and the line counts, by the first crosslink --lines. The line counts are the whole upgrade cost, and they grow with a store's history, not its size: stores with long histories take longest. Measured on a forge's stores (8 cores, zit at low CPU priority beside the live forge), first command after upgrading from caches 0.12.1 built:

store	size	crosslink --lines	peak memory	info	ls --long tip	annotate -n 50
tcl	420 MB	717 s	218 MB	0.01 s	0.11 s	0.54 s
sqlite	331 MB	366 s	220 MB	0.02 s	0.04 s	0.28 s
pkgsrc (sub-manifests)	632 MB	299 s	781 MB	0.13 s	0.94 s	0.16 s

Later crosslink --lines calls take under a second. A store with no .zit-cache at all builds every cache on its first commands; on the same forge, from an empty cache, crosslink --lines took 1,524 s on netbsd-src (3.8 GB, 928 MB peak), 681 s on tcl, 353 s on sqlite, 341 s on fossil and 275 s on pkgsrc, and under 50 s on every smaller store.

A server can build the new caches ahead of time on a copy of each store, taken with its .zit-cache (crosslink --lines, then ls --long tip), then move the copy's .zit-cache into place at the switch, as for 0.12.0. Do not let an older zit open a store whose caches are in the new format: it rebuilds its own over them.

Merges that both add lines at the same point now conflict where 0.12.1 took both sides; see above.

Qualification

The release manifest records native qualification against the exact tagged source archive on macOS ARM64, Linux x86-64 and Linux ARM64, using the pinned Zig 0.17.0 toolchain. Publication is gated on unit and end-to-end tests, Git differentials, deterministic simulation, seeded fuzzing, and Fossil interoperability checks on every distributed target.

For the basis of the conformance claim, see the 0.1.0 release notes.