ADR 0233: JUnit XML joins diagnostic auto-detection

Status: accepted. Extends ADR 0028's Tier-1 normalization and follows ADR 0083's offload boundary; the declared text-format slice stays future work.

A failed gate job becomes structured diagnostics only when its output arrives in a machine format discern recognizes. Until now that meant SARIF, which linters and type-checkers emit and test runners generally do not. A red test stage therefore produced one Tier-0 diagnostic: the job label, the exit code, and a captured output blob.

The local logbook made the cost concrete. It records diagnostics as classes — tool, rule, file — so a recurring failure can be attributed across branches. With test failures carrying neither rule nor file, the record read test and nothing else. When the practice logbook flagged one tree flipping red and green across four identical runs, no recorded evidence could name the test that flipped: the flake analysis had to be reconstructed from run durations and timestamps.

JUnit XML is the test-runner counterpart of SARIF: a cross-runner report format most test tools can emit behind a flag (--reporter=junit, --junit-xml, and similar), naming each failing case, its file when the runner knows it, and its failure message.

Diagnostic normalization auto-detects JUnit XML beside SARIF. The doctrine is unchanged from ADR 0028: discern recognizes the format, never the tool, so the stack-neutral core stays neutral and a project opts in by making its command emit the format — no configuration surface is added.

Detection requires a real <testsuites>/<testsuite> element with a matching close tag, sliced out of surrounding runner noise. Only cases carrying a failure or error element become diagnostics: file prefers the case's own file attribute, then a path-like classname, then a path-like enclosing suite name — a dotted language identifier is never mistaken for a path. The test's name rides in rule, so a recorded diagnostic class identifies the exact test. A report with no failing cases keeps the Tier-0 raw output: a runner that crashed before the suite still shows its real output rather than an empty normalization.

This repository's own test job now runs with --reporter=junit, dogfooding the tier.

• A red test stage names each broken test with file, line, and message instead of handing back one opaque blob; agents jump to the failure without scraping output.
• Logbook diagnostic classes for test failures become attributable, so recurring-failure and flake findings can name the test rather than the stage.
• Parsing is a bounded, defensive scan, not a full XML parser: a writer that leaves a raw > inside attribute values (real emitters escape it) degrades that case's parse, and any unrecognized document falls back to Tier-0 untouched.
• The gate's captured raw output for this repository's test job is now XML. That trade is deliberate: when normalization succeeds the structured findings replace the blob, and when it cannot, the XML still carries the failure text.

• The declared regex slice ([diagnostics.<name>]) first. Rejected for this need: it requires a new config surface, and test runners already share a structured format the SARIF precedent handles config-free. The regex slice remains the plan for tools with no machine format at all.
• Parsing each runner's human text. Rejected: it binds the core to named tools, which the stack-neutral rule forbids.
• A JUnit file path (--junit-path) read from disk. Rejected: diagnostics normalize from captured output; reading report files would add a second evidence channel with its own lifecycle and staleness questions.