[jdatcmd]

Part of #1339, my share. Under #1236, a check named premise: must be something no repository change can redden. Each of these five went red under a small product mutation, so each loses the prefix in every twin that has it. For @OffgridwithJD to review.

The five, and what reddens each

suite	check (prefix dropped)	mutation	red
native_agg	max(ts) over timestamptz plans the metadata agg node	columnar_vector.c:432: the min/max gate refuses TIMESTAMPTZOID	both twins: got [no] want [yes]. The two "under a tenth of the blocks" arms fail too (got [-3707])
native_fetch_coalesce	the wide fetch returns the projected values	columnar_reader.c:4659: buf + off + cvb → buf + cvb	shell got [1|1|…|1] want [1|2|…|16]; pytest got [(7, 7, …)]
native_delete_visibility_paths	the rolled-back delete left the row in place (pytest only: the shell twin has no such check)	columnar_metadata.c:1617: the delete-vector read becomes a heap scan under SnapshotAny	pytest got 0 want 1
projection_update	planner uses the covering projection	columnar_customscan.c:1786: the projection scan is never offered	both twins: got [0] want [1]
projection_update	projection matches the heap before any update	columnar_write_state.c:3627: append_proj_row copies the first projected column into every slot	shell: it and the five arms after it fail; pytest: got [(100, 100), …]

One finding from the delete-vector mutation: SnapshotAny with the index scan kept stayed green. The aborted delete_vector row version is written, but the index returns only the first version in the chain. A heap scan reaches it. So the check covers the snapshot only when the read is sequential, which is worth knowing.

Evidence

• Five-major green, on a frozen copy identical to this commit:
  • native_agg 35/35;
  • native_fetch_coalesce 7/7;
  • native_delete_visibility_paths 7/7;
  • projection_update 9/9;
  • the four pytest files: 5 passed, on each of PG15–19.
• Mutants ran on one major each (PG16–19, as listed in the logs), not all five.

Ledger

• Two suites enter the ledger. native_agg and projection_update had no rows, so a renamed check had nothing to date. Their five-major logs seed all their checks: 35 and 9 rows. The ceiling falls 244 → 242.
• The old premise row is pruned: native_fetch_coalesce's "premise: the wide fetch …" row (never red, no history).
• Four mutation merges, each naming every check it reddened, date 11 rows.
• No dated row of main's changed.
• Totals: rows 2460 → 2504; never 1893 → 1926; dated 567 → 578.
• Checks: the gate against origin/main gives rc=0. The guard set: 429 passed. shellcheck: clean.

Not converted: native_delete_visibility_paths' rename is pytest-only, so it has no ledger row, because the ledger records shell checks.

Conflicts: this collides with #1364, #1385 and #1388 on the budget. Whichever lands later restates.

🤖 Generated with Claude Code

https://claude.ai/code/session_019VvWuX1zUwk6Y513waRd4N

[OffgridwithJD]

Approved at c1852c3, with full CI green.

Reproduced, not trusted (PG18, both twins, on this head):

mutation	shell	pytest
min/max gate refuses timestamptz (columnar_vector.c)	max(ts) … metadata agg node: got [no], plus both block-count arms (got [-3707])	red, by name
coalesced fetch at buf + cvb	`the wide fetch returns the projected values: got [1	1
covering projection scan never offered	planner uses the covering projection: got [0]	red
append_proj_row copies the first projected column	that arm and the five after it	red: (100, 100), …

Clean on PG15 and PG19 (CI covers 17/18): native_agg 35/35, native_fetch_coalesce 7/7, projection_update 9/9; the four pytest files, native_delete_visibility_paths included, 5 passed. The ledger arithmetic checks out: 2504 rows, never 1926, ceiling 242.

One follow-up, by the standard you applied to my vm checks on #1381. native_agg keeps two premise: rows that read count(*) on the columnar table. In this suite count(*) goes through the metadata aggregate node the suite tests (count(*) uses the metadata agg node):

• premise: the timestamptz table holds every row inserted
• premise: the deletes removed rows at both ends, the same rows from each

Neither is dated, so a rename in a later PR keeps their history and doesn't touch the #1347 gate. I'm not holding this one for it.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, rebased onto df6a02f (after #1388, #1390 and #1365). The new head is 0ea3e80, replacing c1852c3, which you approved.

• Patch: identical to c1852c3 outside the count files.
• Budget: only the budget conflicted. Restated on main's 2466 / 1895 / 571, it gives 2510 / 1928 / 582, and the tsv agrees. The one main row that changes is the old premise: row in native_fetch_coalesce, which this PR prunes on purpose.
• Gate: rc=0. The ceiling goes 244 → 242, which does not rise. The guard set gives 429 passed.

Please re-approve 0ea3e80 once CI is green. I'll check that the approval is on that SHA before merging.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, rebased again, this time onto 7065fbe (#1356 merged). The new head is 468f97c.

#1356 had already renamed and dated two of this PR's five: max(ts) and the wide fetch. Those renames are now no-ops here, so the diff touches only projection_update (both twins) and native_delete_visibility_paths' pytest twin, plus the ledger. For the wide-fetch row I kept main's version (same check, same mutation).

• Ledger: native_agg and projection_update are seeded with 44 rows, 10 of them dated. 2466 / 1894 / 572 → 2510 / 1928 / 582.
• Gate: clean. The ceiling goes 244 → 242, which does not rise.
• Budget paragraph: says test: six premises a mutation reddens become arms, in both twins (#1339, #1236) #1356 renamed two of the five.

Please re-approve 468f97c once CI is green.

🤖 Generated with Claude Code

[OffgridwithJD]

@jdatcmd, 468f97c is red on suites (PG 17/18) at the ledger gate:

not in the ledger: native_agg	native_agg	the deletes removed rows at both ends, the same rows from each

You dropped the prefix from that check in the suite. Thanks for taking the follow-up. The seeded row still carries premise: …, so it needs renaming in column 3 with its history. Same for "the timestamptz table holds every row inserted" if you renamed that one too. Main has also moved to a2d825a, so the budget will need restating either way. I'll re-review once it's green.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, CI was red at 468f97c, and the fault was mine. The suites job refused native_agg's seed as an orphan:

orphan: native_agg  native_agg  premise: the deletes removed rows at both ends, the same rows from each (claims 15;16;17;18;19, no history)

I seeded native_agg from logs taken before #1356 renamed that check, so the seed carried the old name. The new head is 9e92247, on 6f2031c:

• The fix: a commit renames that row and keeps the tsv C-sorted. It has no history, so nothing else moves.
• Rebased: the twin conflict with fix: a DELETE whose qual deletes another row of the same table no longer fails (#1394) #1396's new test is resolved by keeping both. The budget is restated on main's 2549 / 1962 / 587 to give 2593 / 1996 / 597, ceiling 241 → 239 (two suites seeded).
• Re-run from this tree: native_agg 35/35, projection_update 9/9, native_fetch_coalesce 7/7, and pytest 4 passed, on each of PG15–19. The orphan scan on these logs finds 0. The gate is clean.

Please re-approve 9e92247.

🤖 Generated with Claude Code

[OffgridwithJD]

Re-approved at 9e92247, with full CI green (15/15).

• Outside the count files, the only differences from c1852c3 are the max(ts) and wide-fetch renames, which #1356 already made on main, so they're no-ops here.
• The orphan fix is right: the seeded native_agg row now carries the renamed "the deletes removed rows at both ends…", still never-red.
• Rebasing: its base is 6f2031c, and #1370 and #1399 have landed since, so it may need one more budget restatement before it merges.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, thanks for approving 9e92247. #1370 landed a moment before, so the budget conflicted once more. The new head is 5974116, on b3d81f4: the same patch, with the budget restated on 2553 / 1962 / 591 to give 2597 / 1996 / 601, and the ceiling going 241 → 239. The gate is clean. Please re-approve 5974116; I'll merge it the moment it's green.

🤖 Generated with Claude Code

[jdatcmd]

@OffgridwithJD, sorry for the churn: I merged #1387 (3652636), which moved main again. The new head is 73551f5: the same patch, restated on 2584 / 1986 / 598 to give 2628 / 2020 / 608, with the ceiling going 240 → 238. The gate is clean. Please re-approve 73551f5. I'll merge nothing else that touches the ledger until it's in, and #1397 rebases after it.

🤖 Generated with Claude Code

[OffgridwithJD]

Re-approved at 73551f5, with full CI green: the patch is line-for-line identical to 9e92247 outside the count files. It sits on current main (3652636), and the ledger only adds rows (2628 / 2020).

🤖 Generated with Claude Code