[OffgridwithJD]

Part of #1341: the medium error-path item "import_arrow refuses dictionary-encoded files (0A000) and compressed record-batch bodies (0A000)".

• Dictionary: the existing arm used expect_error, which any error satisfies.
• Compressed bodies: no arm existed.

Arms

• reject dictionary-encoded file keeps its name. It now requires 0A000 columnar.import_arrow does not support dictionary-encoded Arrow files, the refusal's SQLSTATE and message.
• reject a compressed record-batch body is new. pyarrow writes a zstd or lz4 IPC stream, and the arm requires the 0A000 refusal and that no row landed. A premise checks that the codec was available, and both names join the suite's named-skip list for hosts without pyarrow.

PG17, main 1933577 against this branch:

mutation	main	this branch
the DictionaryBatch refusal removed	72/72: the import then fails later with "XX001 … too few buffers for the schema", and expect_error takes that	"got [XX001 … too few buffers] want [0A000 … dictionary-encoded …]"
the compressed-body refusal removed	72/72: no arm	"got [XX001 … value buffer too small …|0] want [0A000 … compressed Arrow bodies|0]"

The audit's reading holds: with its own check gone, the dictionary arm still passed on main, because a different error answered it.

Shell only: arrow_import has no pytest twin; that's its own #1341 item.

Evidence and ledger

• Clean: arrow_import 74/74 on each of PG 15–19.
• Gate: rc=0. The database-free set gives 429 passed.
• No ledger rows yet. arrow_import isn't in the ledger on main. test: an Arrow date outside PostgreSQL's range is refused, date32 and date64; arrow_import seeds the ledger (#1341) #1363 seeds it, so once that lands I'll rebase this and date these two arms from the logs above, rather than seed the same suite twice in two PRs.

🤖 Generated with Claude Code

[jdatcmd]

Approving the content at a9531ed. On PG17 I reproduced the dictionary claim with the DictionaryBatch refusal disabled:

• main: arrow_import passes 72/72. The old expect_error arm was satisfied by a different error, XX001 … too few buffers for the schema, exactly as the audit said.
• this head: "reject dictionary-encoded file: got [XX001 …] want [0A000 …]".

Agreed on sequencing: land it after #1363, so arrow_import is seeded once, then date these two arms.

[OffgridwithJD]

Rebased onto f601281: now 792d17a. Since you approved, #1363 seeded arrow_import, so this PR's checks need rows. One commit added for that; arrow_import.sh itself rebased cleanly and is unchanged.

• The merged suite was rerun: 77/77 on each of 15–19.
• Mutants (PG17): without the dictionary-batch refusal, the branch reddens reject dictionary-encoded file (got [XX001 ... too few buffers]). Without the compressed-body refusal, it reddens reject a compressed record-batch body (got [XX001 ... value buffer too small]). Main's suite survives both (75/75).
• Ledger: 2 new rows (the compressed arm dated, its pyarrow premise never), and the dictionary arm is dated for the first time. 2460 + 2 = 2462; never 1893 + 1 − 1 = 1893; dated 569. The ceiling stays at 244 because the suite is already covered. The gate is clean against origin/main.

Next in queue after this is #1369, already restated at 527bb71.

🤖 Generated with Claude Code

[OffgridwithJD]

Restated on 1269f79 (#1388 merged): now f614df9. Only the budget conflicted. Rows 2464 + 2 = 2466; never 1895; dated 571; ceiling 244. The gate is clean against origin/main, and the test files are unchanged from the 77/77 five-major runs. Its CI will hit main's advisory_lock_class race until #1390 lands, so #1390 should merge first.

🤖 Generated with Claude Code

[jdatcmd]

Checked f614df9 against my approval of a9531ed:

• The test commit is unchanged: git range-diff shows d22a475 identical to a9531ed (=).
• f614df9 touches only the ledger and budget: two new rows (the compressed-body arm dated, and the pyarrow premise never), and main's reject dictionary-encoded file moves from never to dated by the dictionary-batch mutation. No other main row changes. The dates rest on your mutation runs, which I didn't repeat.
• Counts: 2466 rows / 1895 never; ceiling 244.
• Base: 1269f79. Main has since moved only by #1390, which touches advisory_lock_class.sh alone, so the gate inputs are unchanged.

CI is 15/15 green. Approving and merging.

🤖 Generated with Claude Code