CODE-070: subRevs rescans the root tree per commit — 35% of the git.git walk

Now: OPEN
Sev: MED
Rep: ///bee
Ask: gritzko
Rev: ///bee#8cb69a0
See: BEE-006

Input

Context

index/index.js:271:Wt subRevs costs 25.5 s of a 73.4 s commit walk over git.git (81726 commits) for ONE submodule: subTree misses once per commit (81656 times, 25.3 s) while the ODB reads themselves are 2.65 s — the JS cursor loop re-parses the root tree into a second capped cache. Memoizing subAt and pruning tree === pTree bought nothing (each commit's root tree is distinct). :1125 also takes submodulePaths from the TIP only, so a submodule moved or removed before the tip gets no dir revs, and every gap run re-walks the tip tree for the .gitmodules gate (221 tree parses for a 1-commit catch-up). Found in the 2026-09-18 index review.

Goals

Let git.getTreeDiff keep gitlinks (BEE-006:54:on is why it drops them) or add a leaf answering one tree's gitlink at a path; else readTree keeps gitlinks in its single parse. Gate the .gitmodules check on the tip's cached tree.

WIP

Design decisions

subTree reads the gitlinks off the RAW tree bytes, not the git.tree cursor: that cursor mints a JS object and a 40-hex string PER ENTRY, while only a gitlink's mode starts "16", so two byte compares reject the other 558 entries of git.git's root tree. 3000 root trees: 1510 ms by cursor, 311 ms by scan, the ODB reads inside both some 300 ms.
The tip's sub paths now come from .gitmodules (BEE-006:34:on allows it) with the TREE as the authority: a named path is kept only where subAt finds a real gitlink. A gap run reads one blob instead of parsing all 226 dirs of git.git's tip tree: 6.0 ms -> 1.1 ms per enumeration.
The path list GROWS during the walk — every commit whose diff carries .gitmodules re-reads it — so a sub that moved or was dropped before the tip still gets its revs. One that never had a sub pays nothing: no .gitmodules in any diff, no list, no per-commit probe.

TODOs

[v] bench: git.git walk without the subRevs tax
[v] gitlinks off the one tree parse

Blockers and bummers

The per-commit root tree READ stands: git.getTreeDiff drops gitlinks in dog/git C (BEE-006:54:on), so nothing but a second ODB read answers "the gitlink at P in this tree". The C fix (gitlinks kept in GITTreeDiff, or a leaf for one tree's gitlink at a path) is gritzko's call; it would take the ~8 s of ODB reads this leaves in the walk.
A gitlink .gitmodules does not name is no longer enumerated (the whole-tree walk found it). Git treats such a sub as unconfigured, and the old gate already refused a repo whose tip carries no .gitmodules at all.
A sub added AND dropped inside commits an earlier run already indexed, the tip naming none, still gets no revs: it sees no .gitmodules in any diff. A rebuild does.
bench, bee index over a git.git clone (81726 commits, two runs each, niced): before 92.8 s / 90.5 s, after 69.1 s / 73.4 s, both sides 250344 revs and 940189 rows.