Beagle SCM

CODE-073: MARK rows accumulate — the scan's diff base grows one tree per run

Now

OPEN

Sev

MED

Rep

///bee

Ask

gritzko

Rev

///bee#8cb69a0

See

BEE-063

See

BEE-031

Input

Context

index/lindex.js:201:PS,243-254: a wh128 lane is unkeyed, so every run adds a MARK row and none is removed; base/pTrees therefore grows by one commit per index run and descend pays |pTrees| git.getTreeDiff calls at every tree level, forever. Confirmed: 7 bee index runs over 7 commits left 13 mark rows per ref and a rebuilt step 2 got pTrees = 13 distinct trees. With the BEE-031 post-commit hook that is one extra parent tree per commit — thousands on a live repo. Found in the 2026-09-18 index review.

Goals

Keep only the marks the diff needs — the newest few by CTS, or the single tip whose tree is an ancestor of all — or compact the MARK key space on write.

WIP

Design decisions

• The base is the NEWEST FOUR mark trees by the commit's own date, deduped by tree (index/lindex.js:225:PS baseTrees). A dropped mark loses no row: descend prunes a path equal to ANY base tree, so an older mark only ever saves a re-lex and the mint is idempotent either way.
• Sorted by the COMMIT's own committer date, not by the CTS row: the mark commit is read for its tree in any case, and a lane whose walk never reached it still sorts.
• Four, not one: a branch switch leaves marks on two tips and the newest alone would re-walk the other's tree every run.
• rec.base names the tree count — what the repro pins; the scanned ... summary line is unchanged.

TODOs

• [v] repro: N runs, N marks, N pTrees
• [v] mark set bounded

Blockers and bummers

• The mark ROWS still grow one per run: a wh128 lane has no delete and _compact_wh128 merges runs, never drops one. Compacting the MARK key space wants a keyed lane (kv64) or a C-side delete — gritzko's call. The cost is 16 B and one seek per run; only the diff base ever paid more.
• index/index.js:452:Wt markSet grows the same way for the ref watermark, but it is a membership test and no diff base, so it was left alone (another agent holds that file).

Outcome