fix(sql): recover T-SQL bracket-named, CREATE OR ALTER, and PROC routines

[egarcia74]

Problem

T-SQL stored procedures and functions are silently dropped from the graph. On a real 391-file T-SQL data-migration corpus (SQL Server DACPAC project: staging/model/transform layers, [Schema].[Object] naming throughout), 0 of 26 stored procedures produced a node — while tables and views extracted fine, so the gap is invisible unless you go counting.

Two independent defects stack:

1. The grammar can't parse the T-SQL routine shape at all. CREATE PROCEDURE ... AS BEGIN ... END never parses structurally in tree-sitter-sql (there is no create_procedure production for the AS BEGIN body idiom), so ERROR-node regex recovery is these objects' only path into the graph.
2. The recovery regex doesn't speak T-SQL. It accepted only unquoted or double-quoted names — no bracket-delimited names (CREATE PROCEDURE [dbo].[usp_Load]), no CREATE OR ALTER, no PROC shorthand. On this corpus, every routine uses brackets, so recovery recovered nothing.

What this PR does

All changes are in graphify/extractors/sql.py:

• One shared pattern, _ROUTINE_RECOVERY_RX, hoisted to module level and used by both recovery sites (the walk-time ERROR-node scan and the whole-file has_error fallback). The two sites had drifted; the drifted second site is what minted a phantom dbo() node for a mixed-delimiter name (CREATE PROCEDURE dbo.[usp_Mixed]) — the schema captured as if it were the routine. Sharing the pattern fixes that and prevents re-drift.
• Bracket-delimited identifiers, including T-SQL's ]] escape for a literal ] inside a name ([dbo].[a]]b] recovers as one routine, fully consumed, not truncated at the escape).
• CREATE OR ALTER alongside the existing OR REPLACE, and the PROC shorthand alongside PROCEDURE.
• Comment masking in the whole-file fallback: the raw-text scan now runs over _mask_sql_comments(src) (offset-preserving blanking of -- and /* */), so commented-out DDL in a file with an unrelated parse error can no longer fabricate routine nodes.

Recovered routines are name-only nodes (no body reads_from edges), matching the existing PL/pgSQL recovery behavior — a named node with a contains edge beats an invisible object.

Result on the real corpus

391 .sql files, before → after: 1 → 28 routine nodes, including every bracket-named [Audit]/[Utils]/[Monitoring] procedure; table and view counts unchanged; no phantom schema nodes.

Tests

Added in tests/test_multilang.py, each verified to fail with the fix reverted:

• bracket-delimited procedure recovers under its full [dbo].[usp_X] name;
• CREATE OR ALTER PROCEDURE recovers;
• PROC shorthand recovers;
• mixed-delimiter name (dbo.[usp_Mixed]) yields exactly one node — both recovery sites agree on the captured name (the phantom-dbo() regression);
• ]] escape inside a bracket-delimited name is consumed, capturing [dbo].[a]]b]() as one routine;
• commented-out DDL in an ERROR-bearing file does not fabricate a routine node.

Validation

• Full suite: 6 failed, 4836 passed, 28 skipped — the 6 (test_ollama.py, test_ollama_retry_cap.py) fail identically on unmodified v8 in this environment (local Ollama config), none touch SQL.
• ruff check graphify/ tests/: all checks passed.

Note

Independent of my companion PR #2959 making tree-sitter-sql a core dependency; both touch tests/test_multilang.py, and a local merge of the two branches resolves trivially (the combined branch produced the 1 → 28 measurement above and runs in production).

[graphify-labs]

Graphify reviewed this change.

Worth a look — the grounded gate found no coupling regressions or blocking issues, but 1 advisory finding(s) below merit a look before merge.

Formal verification. No changes could be formally verified in this run.

---

Graphify review — findings

Extends SQL routine recovery to handle T-SQL dialect: adds bracket-delimited identifiers ([dbo].[usp_Load], with ]] escaping), CREATE OR ALTER, and the PROC shorthand, and consolidates the previously-duplicated recovery regex into a shared _ROUTINE_RECOVERY_RX so the walk-time ERROR scan and whole-file fallback can't drift and double-emit. Adds _mask_sql_comments (offset-preserving) around the whole-file scan so commented-out DDL can't fabricate routine nodes when an unrelated parse error arms recovery. Covers all of this with new test_sql_* cases.

Worth a look

• Comment masking treats -- inside string literals as a real comment — graphify/extractors/sql.py:37 · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 279 functions depend on the 117 functions this change touches.

Health — this change adds coupling hotspots:

• new: dispatch_command() — 2 callers, 119 callees
• new: extract_sql() — 15 callers, 9 callees
• new: walk() — 1 callers, 8 callees
• new: test_poisoned_manifest_is_healed() — 0 callers, 6 callees

Verification — 279 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 128 function(s) in the blast radius were not formally verified this run

Formal verification

Could not verify: Could not verify extract\_sql.

The verifier did not have enough to check extract\_sql, so it is saying so rather than guessing. No false assurance is the whole point.

Guarantee: No guarantee either way, this is an honest abstention, not a pass.

Note: Reason: parameter `path` is annotated `Path` — outside the synthesizable primitive/collection set

· 4 more finding(s) on lines outside this diff (see the check run).

[egarcia74]

Addressed the Graphify-review finding (-- inside string literals treated as a comment opener) in b6d4dfc, and it turned out to be half of a symmetric pair:

• The mask is now literal-aware: single-quoted strings (with '' escapes), double-quoted identifiers, and bracket-delimited identifiers (with ]] escapes) are preserved before comments are blanked, so '-- note' / 'open /* here' can no longer hide DDL sharing the span. Literal patterns are deliberately single-line — this mask only runs on files that already failed to parse, where an unclosed quote is likely, and a multi-line match would let one unclosed delimiter swallow real DDL below it (a miss is visible; a swallow is silent).
• The walk-time ERROR-node scan now masks too: while verifying the fix I found the fabrication risk the whole-file scan was cured of was still live at the other recovery site — an ERROR blob whose byte span covers commented-out DDL minted a routine node from it (reproduced with a -- CREATE PROC ... line sandwiched between broken segments).

Both directions are mutation-tested: a literal-blind mask fails the new unit pin (test_mask_sql_comments_preserves_literals_and_blanks_comments), and an unmasked ERROR scan fails the extended fabrication test. Full suite after the change: 6 failed, 4838 passed, 28 skipped — same 6 environmental Ollama failures as on unmodified v8; ruff check clean.

[graphify-labs]

Graphify reviewed this change.

Worth a look — the grounded gate found no coupling regressions or blocking issues, but 2 advisory finding(s) below merit a look before merge.

Formal verification. No changes could be formally verified in this run.

---

Graphify review — findings

Extends SQL routine recovery in extract_sql to handle T-SQL bracket-delimited identifiers ([dbo].[usp_Load], with ]] escapes), CREATE OR ALTER, and the PROC shorthand, and consolidates the walk-time ERROR scan and whole-file fallback onto a single shared _ROUTINE_RECOVERY_RX so the two sites can no longer capture divergent names and double-emit nodes. Adds _mask_sql_comments to blank comment spans (offset-preserving, literals kept verbatim) at both recovery sites so commented-out DDL can't fabricate routine nodes. Adds tests covering the bracketed/OR ALTER/PROC forms, escaped-bracket names, cross-site name agreement, and comment masking.

Worth a look

• Routine recovery still matches DDL text inside string literals — graphify/extractors/sql.py:510 · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review
• Optional tree_sitter_sql dependency is imported instead of skipped — tests/test_multilang.py:536 · Escalate · medium
  • agreed by 2 of 2 members but NOT verified (no proof, no reproducing execution) — consensus is not a verdict; needs human review

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 283 functions depend on the 121 functions this change touches.

Health — this change adds coupling hotspots:

• new: dispatch_command() — 2 callers, 119 callees
• new: extract_sql() — 16 callers, 9 callees
• new: walk() — 1 callers, 9 callees
• new: test_poisoned_manifest_is_healed() — 0 callers, 6 callees

Verification — 283 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 132 function(s) in the blast radius were not formally verified this run

Formal verification

Could not verify: Could not verify extract\_sql.

The verifier did not have enough to check extract\_sql, so it is saying so rather than guessing. No false assurance is the whole point.

Guarantee: No guarantee either way, this is an honest abstention, not a pass.

Note: Reason: parameter `path` is annotated `Path` — outside the synthesizable primitive/collection set

· 4 more finding(s) on lines outside this diff (see the check run).