The canonical-form and arm-order checkers run in CI, gated on their findings

[Rafael-SOWNet]

Part of #1256: the last two of its self-contained harnesses.

CanonCheck and Confluence report findings by design. InnerSimplified is a normalisation and does not reorder operands, so 2024 of its 2738 order checks fail. Three pairs of arms in Common and Power disagree where both fire. A gate that fails on a non-empty list would always be red.

So each commits its list of findings beside it, the way corpus-baseline.tsv does, and fails when the list changes in either direction. The list is the gate, and the counts can move freely.

harness	list	a line is	run
CanonCheck	2091	stage, property, and the case written out in full	1.6 s
Confluence	3	the set and its two arms, by rule name	0.8 s

Arms are named by their rule names, such as a-variable-times-a-number-puts-the-number-first, rather than by index. A name survives an arm being inserted above it.

• HARNESS_UPDATE_BASELINE=1 rewrites a list. The failure message says so, and lists what appeared and what went away.
• A finding whose case timed out counts as unmeasured, not gone. A timeout still fails nothing.
• The shared code is Shared/Baseline.cs.

Checked:

• Each list is the same across three processes.
• File-level control, both harnesses, both directions. I removed one committed line and the run exited 1 with "1 appeared". I added a line that cannot occur and the run exited 1 with "1 went away". With the list restored, both exit 0.
• Library-level control, Confluence only. I made a-variable-times-a-number-puts-the-number-first hand its node back unchanged. Confluence exited 1 and named the Common pair that went away. CanonCheck's list was unchanged, because no finding of its involves that arm. Reverted.
• Not exercised: the timeout path. No case timed out in any run.

AGENTS.md's two mentions of the harnesses in work/ now point to Sources/Tests/Harnesses. So do five code comments and CanonicalForm.md. Its section 7 had listed three candidate forms where the harness measures four. The catalogue's count of idempotence failures for the sort alone was 21, and it is 18 on master. BREAKING-CHANGES.md keeps its two mentions, which record where a change was found at the time.

After this, #1256's remaining harnesses each need something outside the repository:

• docsamples needs the wiki and the site;
• intbench needs the Rubi suite;
• libcompare needs two other libraries from NuGet;
• egraph is a measurement rather than a check.

🤖 Generated with Claude Code

https://claude.ai/code/session_012sonx8iAspMiwRwokT1Ura