Skip to content

fix(tri): marker before dirty-guard; and a control that kills nothing (#2492) - #2535

Merged
gHashTag merged 2 commits into
masterfrom
fix/marker-before-dirty
Aug 23, 2026
Merged

gHashTag merged 2 commits into
masterfrom
fix/marker-before-dirty

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

The interrupt marker caught a real one

A run from the previous iteration had been orphaned and left gft_backprop_microcode.py mutated. The marker named the gate, printed the recovery commands, and I ran exactly what it printed — git checkout -- tools/ on the directory, which is the instruction #2530 says I should have followed the first time.

And it spoke second, which is the wrong order. The dirty-tree guard fires first, and after an interrupt the tree is dirty — so the informative message existed and was never the one shown. The marker is now checked first. Found by hitting a real interrupt and reading the wrong error.

An orphaned run was still alive

Thirteen minutes in, mutating files, from an iteration I had already reported as finished. Third time mutants got loose, and every time the cause was mine: starting a long background job and losing track of it.

The guard did refuse the concurrent run I started on top of it — an unplanned safety property, since two mutation loops would each restore the other's mutations and produce numbers that describe nothing.

Seventeen of twenty-one rows, and one is stark

verify_emit_bitexact.py   0/1   0/1   0/11   0/4   0/0

Its control kills nothing. Not one of seventeen mutants, across four operators.

I gave that gate a control three iterations ago (#2520) — three planted cases, all passing, covering the skip pair and the codegen-failure branch — and none of the sites the operators can reach are among them.

Same class as verify_exhaustive.py before #2515: a control that exists, passes, and is not connected to the verdicts anyone would break. Worse here, because I wrote this one knowing that class, and the number that exposes it could not be produced until the run learned to be interruptible (#2533).

The table is 17 of 21 and says so

The remaining four are still measuring; the cache means the next run finishes them rather than starting over. A partial table reported as partial is a measurement; the same table reported as complete is the thing this campaign is about.

Refs #2492

@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

This notebook contains session context, decisions, and artifacts for this work.

@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-23 11:56:22 UTC

Summary

Status Count
Total Open PRs 7
PRs with Failing Checks 3
PRs with All Checks Green 4
READY 3
FAILING 3
PENDING 0

Seal Status

  • ⚠️ STALE -- sha256(compiler.rs)=e905ba8eeb53 != manifest seal=87e5cbd3ad94.
    The committed NMSE numbers were certified against an older compiler.rs.
    Run scripts/reseal-check.sh locally for the two-step reseal command (advisory; not a merge gate).

@gHashTag
gHashTag merged commit e1dd18e into master Aug 23, 2026
31 of 33 checks passed
@gHashTag
gHashTag deleted the fix/marker-before-dirty branch August 23, 2026 11:58
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant