Problem

Both RTS helpers open a fresh WT_CURSOR_BTREE, call a search function to locate the row/record, then apply the update — even though the caller already holds the exact page and slot:

• _rts_btree_row_modify (src/rollback_to_stable/rts_btree.c:274) calls _wt_row_search with the key, but both call sites (lines 629, 836) already hold WT_REF *ref and WT_ROW *rip — the exact page and row slot.

• _rts_btree_col_modify (src/rollback_to_stable/rts_btree.c:244) calls wt_col_search with the recno, but the column-store iteration loop at line 860 already holds WT_COL *cip (and thus the slot via WT_COL_SLOT(page, cip)) before dispatching to _rts_btree_abort_ondisk_kv.

In both cases the search is redundant: the position is already known before the function is called.

Proposed Optimization

Thread the known slot pointer into each helper (or position the cursor directly) instead of searching:

Row store — pass WT_ROW *rip and set the cursor position directly:

cbt.ref     = ref;
cbt.slot    = WT_ROW_SLOT(ref->page, rip);
cbt.compare = 0;
cbt.ins     = NULL;

Column store — pass WT_COL *cip down through __rts_btree_abort_ondisk_kv and set:

cbt.ref     = ref;
cbt.slot    = WT_COL_SLOT(ref->page, cip);
cbt.compare = 0;
cbt.ins     = NULL;

This lets _wt_row_modify / _wt_col_modify take the existing-slot update path (cbt->compare == 0, cbt->ins == NULL) without any tree traversal or binary search.

Impact

RTS iterates over every key on every dirty page during startup recovery and prepared-transaction rollback. Each call to these helpers currently performs a redundant B-tree search even though the caller already has the slot. Eliminating both searches reduces CPU overhead on the critical recovery path.

Acceptance Criteria

• _rts_btree_row_modify no longer calls _wt_row_search when rip is available
• _rts_btree_col_modify no longer calls _wt_col_search when cip is available
• Existing RTS csuite / Python suite tests continue to pass