Beagle SCM

CODE-056: repsOf: the WALK_CAP truncates silently, once per path per tip

Now

OPEN

Sev

MED

Rep

///bee

Ask

gritzko

Rev

///bee#8cb69a0

See

BEE-005

See

CODE-068

Input

Context

index/dag.js:103:~c with :12 WALK_CAP (65536) stops the ancestor walk and says nothing. On /src/linux drivers/gpu/drm/i915/i915_drv.c (deleted at HEAD, 1458 revs) gets 0 reps at the default cap and 27 at 400k: zero reps drops the path into the degraded empty-seed fold, and a PARTIAL truncation is worse — a wrong floor and spurious hunks with no message. The walk is also redone per path per tip with no parentsOf memo (375 ms per capped walk on linux, 75 s before any folding for a 100-path diff); view/status.js:223:vX,379 builds full ancestors sets through the same cap. Found in the 2026-09-18 review.

Goals

One commit-DAG walk per diff serving every path, stopping at the frontier of found reps rather than at a global commit count, and a loud note when a cap does bite.

WIP

TODOs

• [x] repro: deep deleted path answers 0 reps under the cap
• [x] shared walk per diff; cap reported

Design decisions

1. The reps walk gets a bound of its own, REP_CAP (2^21), apart from WALK_CAP: it stops at the FRONTIER of the revs found, so the bound is a brake and not the rule. Only a commit the walk EXPANDS spends it, and past it the queue is still read for revs: a bite costs depth, never a rev already in hand.
2. A bite is io.log'd and left on the result as capped, so the empty-seed fold in index/weave.js:203:KB is never silent again.
3. parentsOf memoises its CPAR rows per index handle (a WeakMap), so one diff walks the commit dag once for every path it folds.
4. repsOf answers the MAXIMAL revs it found, not every one: a rev another found rev descends from already rides in that one's closure, so it stands for nothing of its own. On linux that is 345 -> 28 reps for i915_drv.c, 4045 -> 107 for MAINTAINERS. CODE-068 plans the same reduction inside floorRev; with this one in repsOf, that half of CODE-068 has nothing left to do.

Blockers and bummers

• The complete walk for a path deleted 417743 commits ago costs 8.1 s and 673 MB RSS on ~/src/linux, paid once per diff (the next two paths cost 2.3 s and 1.1 s off the memo). That is the price of an answer where there was none at all, and REP_CAP is the one constant to turn if the trade is wrong.
• view/status.js:223:vX only NOTES its cap: ahead/behind still truncate at WALK_CAP, since a full ancestor set is what the divergence algorithm itself wants — its own ticket, not this.

Outcome