Greptile Summary

This PR fixes two related issues with schema-qualifier normalisation for PostgreSQL functions and procedures:

1. Issue plan on same db that was dumped isn't empty #360 — stripSchemaFromReturnType was returning TABLE(...) return types unchanged (early-return with "already handled" comment) even though the TABLE column types could be schema-qualified by pg_get_function_result. The fix adds proper column-level parsing to strip the schema prefix from each column's type.

2. Issue SET search_path not showing correct on plan or apply if you use '' #354 follow-up — Schema qualifiers inside dollar-quoted function/procedure bodies are now preserved verbatim instead of being stripped at normalisation time. A new splitDollarQuotedSegments helper skips $$…$$ / $tag$…$tag$ blocks, and the diff comparison uses a new definitionsEqualIgnoringSchema helper that strips qualifiers at comparison time only.

The changes are well-structured and cover the main regression path. Two concerns worth addressing:

• Fragile comma-split for parameterized TABLE column types: Both the existing normalizeFunctionReturnType and the new stripSchemaFromReturnType TABLE handling use strings.Split(inner, ","). This will mis-parse column types that contain commas in their precision/scale specification (e.g., numeric(10, 2)), silently producing malformed column lists. A parenthesis-aware split should be used in both functions.
• Test coverage gap: The integration test for issue plan on same db that was dumped isn't empty #360 uses identical old.sql / new.sql files, validating idempotency but not the specific schema-stripping path. A fixture where one side has public.datetimeoffset and the other has datetimeoffset would directly exercise the fix.
• Quoting inconsistency in golden file: testdata/dump/issue_320_plpgsql_reserved_keyword_type/pgschema.sql now shows public."user" (quoted) in one function but public.user (unquoted) in another, for the same reserved-keyword table name. This likely reflects genuine pg_get_functiondef output, but should be verified to ensure it round-trips safely.

Confidence Score: 3/5

• This PR is generally safe to merge but contains a logic issue with comma-splitting of parameterized TABLE column types that could silently corrupt return-type comparisons for affected schemas.
• The core fix for issue plan on same db that was dumped isn't empty #360 is correct and well-tested for simple custom types. The dollar-quote preservation logic for function bodies is well-implemented with good unit tests. However, the strings.Split(inner, ",") used in both stripSchemaFromReturnType and the pre-existing normalizeFunctionReturnType will produce incorrect results for TABLE return types containing parameterized column types (e.g., numeric(10, 2)), which is a P1 logic bug. Additionally, the test fixture for the primary fix only validates idempotency rather than the schema-stripping path itself, reducing confidence that the exact scenario from the bug report is covered.
• ir/normalize.go — the TABLE comma-split logic in both normalizeFunctionReturnType and stripSchemaFromReturnType; testdata/dump/issue_320_plpgsql_reserved_keyword_type/pgschema.sql — inconsistent quoting of the user keyword.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[ReturnType from pg_get_function_result] --> B{SETOF prefix?}
    B -- Yes --> C[Strip schema from SETOF base type]
    B -- No --> D{TABLE prefix and suffix?}
    D -- Yes --> E["Split inner by comma (strings.Split)"]
    E --> F[Strip schema from each column type]
    F --> G["Rejoin → TABLE(col type, ...)"]
    D -- No --> H[Strip schema from direct type name]

    I[Function Body from pg_get_functiondef] --> J[splitDollarQuotedSegments]
    J --> K{Inside dollar-quoted block?}
    K -- Yes --> L[Preserve verbatim]
    K -- No --> M[stripSchemaQualificationsFromText]
    M --> N[Apply regex patterns 1-4]

    O[Diff comparison] --> P[definitionsEqualIgnoringSchema]
    P --> Q[StripSchemaPrefixFromBody on both sides]
    Q --> R{Stripped forms equal?}
    R -- Yes --> S[No diff generated]
    R -- No --> T[CREATE OR REPLACE or DROP + CREATE]

_{Last reviewed commit: "fix: strip schema qu..."}