[PATCH v2] ref-filter: restore prefix-scoped iteration

Commit dabecb9db2 (for-each-ref: introduce a '--start-after' option, 2025-07-15) changed single-kind branch, remote-tracking branch, and tag enumeration in do_filter_refs() from constructing an iterator with the namespace prefix to constructing an unscoped iterator and applying the prefix with ref_iterator_seek().

Before that change, refs_for_each_fullref_in() passed the namespace prefix during iterator construction. That helper has since been replaced by refs_for_each_ref_ext().

The files backend primes its loose-ref cache for the construction prefix before it opens packed refs. An empty construction prefix therefore reads every loose ref, and a later seek cannot undo that I/O. Consequently, git branch, git branch --remotes, and git tag scale with unrelated loose refs.

Patrick Steinhardt observed during review that iterator construction and seeking accepted similar strings but assigned them different state semantics. Junio C Hamano then pointed out that no current command can combine start_after with this single-kind path, but future branch or tag support would need to keep the namespace while moving the cursor.

Keep the existing start_after path unchanged. The iterator API cannot currently seek to one string while retaining another as its prefix: an unflagged seek clears the prefix, while REF_ITERATOR_SEEK_SET_PREFIX replaces it with the seek string.

For the commands affected by this regression, which do not set start_after, pass the namespace prefix during iterator construction so that loose refs are scoped before the packed-refs snapshot is opened. This fixes the current regression without deleting the ref-filter state discussed during review or changing its dormant behavior.

Add REFFILES-gated performance cases with one branch, one remote-tracking branch, one tag, and 10,000 unrelated loose refs. The benchmarks were run with:

    GIT_PERF_REPEAT_COUNT=5 GIT_PERF_MAKE_OPTS=-j8 \
        t/perf/run a89346e34a . -- p6300-for-each-ref.sh

The following are the best of five runs, with each run invoking the command ten times. Times are elapsed seconds with user and system CPU seconds in parentheses:

                                  a89346e34a       this commit
  branch                       2.74(0.13+2.56)   0.11(0.04+0.04)
  branch --remotes             2.81(0.13+2.62)   0.12(0.04+0.04)
  tag                          3.01(0.14+2.82)   0.11(0.04+0.04)

Both revisions used the default -O2 build flags and a config.mak containing only "NO_REGEX = NeedsStartEnd". They were built with Apple clang 21.0.0 on macOS 26.5. The machine was a MacBook Pro (Mac16,6) with a 16-core Apple M4 Max (12 performance and four efficiency cores) and 128 GB RAM.

Link: https://lore.kernel.org/git/aGZidwwlToWThkn8@pks.im/
Link: https://lore.kernel.org/git/xmqqikjq7s16.fsf@gitster.g/
Fixes: dabecb9db2b2 ("for-each-ref: introduce a '--start-after' option")
Assisted-by: Codex gpt-5.5
Signed-off-by: Tamir Duberstein <tamird@gmail.com>
---
The series is based on a89346e34a (maint) because the regression has
been present in released versions since Git 2.51.0.
---
Changes in v2:
- Extract local variable `store`.
- Link to v1: https://patch.msgid.link/20260605-fix-git-branch-regression-v1-1-02f40ad40929@gmail.com
---
 ref-filter.c                 | 28 +++++++++++++++++++---------
 t/perf/p6300-for-each-ref.sh | 39 ++++++++++++++++++++++++++++++++++++++-
 2 files changed, 57 insertions(+), 10 deletions(-)