Gate run 2857 on 36a73761 ran 3,762 tests with one failure, the report
snapshot database test, so it is the last thing between staging and a deploy.
The brief told the executor to move the shared psql and psqlAs helpers to
stdin. That instruction is withdrawn. The two helpers serve 522 calls across
29 test files, and psql -c runs a multi-statement string as one implicit
transaction. selectAsAuthenticated relies on that: it sets the JWT subject
with set_config(..., true), which lasts only for the current transaction.
Statement-at-a-time execution would run every later RLS query with no user
set, so those tests would stop testing what they claim to test. Five other
files also write their own begin, commit or rollback.
The brief now asks for a separate single-transaction stdin helper with an
explicit maxBuffer, used only by the snapshot test, with a test proving a
transaction-local set_config stays visible. It also says that if the snapshot
test then fails a real assertion, that is a product defect to report, not an
assertion to relax.
Privacy scan on this tree before push: zero findings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
Acceptance of a12f2c0d against TASK-report-density-fix-20260923: F1 to F6
all pass. On the fixture the fact tables are 130 rows across 600 visible
cells with no engine key, no array index and no long decimal, and every
sourcePath resolves back into the packet. The first major period reads 18
years. Python and TS appendix cleaning are byte-identical. Removing the
writer table guard turns four of its six tests red. Under Node 22 the
frontend goes from 3,730 tests / 26 failing to 3,748 / 26 with an identical
list; the Python quick gate goes from 946 passed / 1 failed to 948 / 0.
One thing still fails: the database test for the report snapshot passes a
428,557-byte SQL string to psql as a single argument, past Linux's 131,072
byte limit, so it dies with E2BIG before reaching Postgres. The snapshot row
has never run against a real database. Recorded as BUG-1011 with a fix brief.
The real-person checklist the executor's tool refused to write is added here.
Privacy scan on this tree before push: zero findings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017eEAG8HD3mm8gsKXgk8uU8
