The deadlock pattern only offers a report whose ERROR: is the line's own label, and the managed prefix's gap can't slide past the real pid (#4014)

[erikdarlingdata]

Closes #4014.

Why

#4014 reported that a backend's STATEMENT companion echoing ERROR: deadlock detected, with a tab-continued DETAIL:, matched the deadlock pattern through the line's real pid bracket. That would attribute a forged deadlock to a real backend. The managed-prefix branch also kept the lazy [^\n]*? gap that #4008 removed from plan capture.

Measured end to end, the forgery never stores anything. Since #4005 every candidate goes through PgLogEntryAssembler (FromReport), whose label is the line's own, and it refuses both shapes. The new parser tests pin that. The pattern itself was still wider than that rule: on PostgreSQL 18.6, the old SQL pattern really did offer the STATEMENT echo as a candidate.

What changes

The C# route (PgDeadlockLogParser.s_deadlockBlock) and the self-hosted SQL route (PgDeadlocksCollector) change together, as defence in depth that matches the assembler's rule:

• The gaps before ERROR: and before DETAIL: may not cross a field label's ": ", so the ERROR: matched is the line's first label. The %Q query id glued to the label still fits, which the captured fixture pins. A prefix that itself renders ": " (an application_name under %a) would hide that line's report; it can hide one, never forge one, and that's documented.
• The managed family's gap to the pid bracket is [^\[\n]*, Plan capture reads a plan block anywhere in a log line, so a statement's author can attach a forged plan to any query id #4008's fix for plan capture, so it can't backtrack past the real bracket.

Not here, filed as #4041: the assembler, and therefore log events and deadlocks, drops every line under a prefix with fields between the zone and the pid. Plan capture has accepted that shape since #4016. Fixing it changes the #3996-hardened reader shared by all log families, so it gets its own PR and review round.

Test plan

• PgDeadlockLogParserTests: new AStatementCompanionEchoingAReport_IsNotReadAsOne (Deadlock capture's regexp_matches may accept a forged STATEMENT echo, same class as #4008 #4014's exact shape, plus a real report after it still reading) and AForgedHeaderBehindARealColonPrefixedBracket_IsNotReadAsOne.
• New PgDeadlockCandidatePatternLiveTests runs the shipped regexp_matches pattern, lifted from BuildQuery, through PostgreSQL's own regex engine:
  • a real report counts as 1 candidate, with or without the glued %Q;
  • the STATEMENT echo counts as 0;
  • pg_deadlocks keeps each deadlock query and the victim statement verbatim, literals included, and returns them to readers and alerts #4005's LOG-statement shape counts as 0;
  • a real report after an echo counts as 1.
• Red-watch: with dev's pattern the live test fails (Expected: 0, Actual: 1, the echo offered).
• Byte pin TheTailerExtraction_LeftBothSiblingsSqlByteIdentical updated for the deadlock SQL.
• Lite: 106/106 (deadlock parser, gate, overlap pin, locale facet). Darling: 117 plus the live class and 26 hygiene gates, green. Builds: 0 warnings.
• Test-only from the user's side: no stored deadlock changes, since the assembler already refused these. No CHANGELOG entry.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GdmA4ND1wLSqA91ax1m4xv