"Every Postgres identifier is schema-qualified" describes 4 DDL files, not the 5 raw-SQL DML sites — and the call-site list is short by 3

[sroussey]

CLAUDE.md states this as an absolute:

Every Postgres identifier is schema-qualified to current_schema() (quote / currentSchemaName in src/util/pgIdentifiers.ts). Unqualified names resolve through search_path and would reach a same-named table in another schema — the hazard applies to DDL, drops, catalog probes and row-count estimates alike.

Measured in the tree, pgIdentifiers is imported by exactly four files, all DDL/catalog:

• src/config/resetAllDatabases.ts
• src/config/alignPostgresColumnTypes.ts
• src/config/addMissingColumns.ts
• src/config/dropStaleCheckConstraints.ts

Every raw-SQL DML site uses bare identifiers:

• src/task/bootstrap/feedFilings.ts:92 — SELECT DISTINCT "filing_date" FROM "filings"…
• src/task/bootstrap/feedFilings.ts:129 — … FROM "filings" WHERE "filing_date" = $1
• src/storage/spac/SpacDealReplace.ts:154 — DELETE FROM "spac_deal" …
• src/storage/form-8k-event/Form8KEventReplace.ts:83 — DELETE FROM "form_8k_events" …
• src/storage/entity/cikNameBulkWriter.ts:110 — INSERT INTO "cik_names" …
• src/task/document/selectFilingsToConvert.ts:250-262 — FROM "filings" f LEFT JOIN "filing_document" d …, plus FROM "spac" s at :246. Added this window.

Nothing is broken today — sec's own deployments run in one schema — but the rule as written is not the rule the code follows, and a new call site landed under it without anyone noticing, which is exactly how a stated-but-unenforced contract stops being useful.

Ask: pick one and make them agree.

1. Qualify the DML sites through quote / currentSchemaName like the DDL tier already does; or
2. Scope the CLAUDE.md rule explicitly to DDL, drops and catalog access, and say why DML is exempt (a single-schema search_path is the deployment assumption).

Second, smaller item in the same paragraph. CLAUDE.md:238-240 enumerates the raw-SQL call sites as:

Call sites: cikNameBulkWriter.ts, Form8KEventReplace.ts, SpacDealReplace.ts (writes); feedFilings.ts (reads).

grep -rn "resolveSqlBackend(" src --include=*.ts | grep -v test returns seven non-test files — the four above plus src/task/spac/spacCandidateScan.ts:230, src/cli/queries/DbStatus.ts:187, and src/task/document/selectFilingsToConvert.ts:177-181. The list should either be complete or be replaced by the grep that produces it.

(For the record, the new site is otherwise careful: selectFilingsToConvert.ts:175-181 gates the fast path on every table the query reads, so one non-durable repo demotes the whole query to the repository path.)

Found during the 2026-08-31 review. Snapshot: workglow-dev/prd → analysis/grades/2026-08-31/sec-detailed.md §4.2.

[sroussey]

Still open, and one site worse. Re-measured at 2ab82ba.

The rule is unchanged in the rewritten CLAUDE.md (.claude/CLAUDE.md, "Cross-cutting rules"):

Every Postgres identifier is schema-qualified to current_schema() (quote / currentSchemaName).

pgIdentifiers is still imported by exactly four files, all DDL/catalog:

$ git grep -ln 'pgIdentifiers' -- src
src/config/addMissingColumns.ts
src/config/alignPostgresColumnTypes.ts
src/config/dropStaleCheckConstraints.ts
src/config/resetAllDatabases.ts

Two of the DML sites this issue listed went downstream with their tiers (SpacDealReplace.ts, Form8KEventReplace.ts). The remaining three are unchanged and a fourth landed this window:

$ git grep -n 'FROM "\|INSERT INTO "\|JOIN "' -- src/task/bootstrap/feedFilings.ts src/storage/entity/cikNameBulkWriter.ts src/task/document/selectFilingsToConvert.ts src/task/facts/factsEligibleCiks.ts
src/task/bootstrap/feedFilings.ts:92:      `SELECT DISTINCT "filing_date" FROM "filings"…`
src/task/bootstrap/feedFilings.ts:129:  'SELECT "accession_number", "cik", "primary_doc", "form" FROM "filings" WHERE "filing_date" = $1'
src/storage/entity/cikNameBulkWriter.ts:110:  `INSERT INTO "cik_names" ("cik", "name") VALUES `…
src/task/document/selectFilingsToConvert.ts:182:  FROM "filings" f
src/task/document/selectFilingsToConvert.ts:183:  LEFT JOIN "filing_document" d
src/task/facts/factsEligibleCiks.ts:85:  `SELECT DISTINCT "cik" FROM "filings"
src/task/facts/factsEligibleCiks.ts:88:   SELECT "cik" FROM "entities" WHERE "sic" IS NOT NULL`

src/task/facts/factsEligibleCiks.ts:85-88 was added in 27280a7 ("Gate the facts eligibility fast path on both tables it reads") — the same pattern as last window's selectFilingsToConvert.ts: a new raw-SQL site landing under a rule stated as absolute, with bare identifiers, and nobody noticing. That is now twice.

The second half of this issue is resolved: the stale call-site enumeration in CLAUDE.md is gone — 5cce549 replaced it with the rule alone, no list. For the record the current count is five non-test call sites (git grep -n 'resolveSqlBackend(' -- src | grep -v '\.test\.' → DbStatus.ts:137, cikNameBulkWriter.ts:42, feedFilings.ts:55,115, selectFilingsToConvert.ts:127-130, factsEligibleCiks.ts:62-64).

Keeping open on the first half. Two cycles.

---

Generated by Claude Code