Skip to content

fix(gates): the last survivor was a frame, not a limit (#2492) - #2502

Merged
gHashTag merged 2 commits into
masterfrom
fix/last-survivor
Aug 23, 2026
Merged

gHashTag merged 2 commits into
masterfrom
fix/last-survivor

Conversation

@gHashTag

Copy link
Copy Markdown
Collaborator

The suite's final surviving mutant was check_elab_ratchet.py's "no baseline" branch, declared UNCOVERED with a reason that had stood for a week:

Reaching it needs t27c fpga-build --smoke to SUCCEED and then find no baseline, and a smoke build that succeeds needs the real spec tree — not something an empty directory can be given. Planting a fake success would test the plant, not the gate.

Every clause is true. The conclusion does not follow. A control does not have to use an empty directory. The note reasoned entirely inside the frame of run_on_empty_tree() — the helper that file happens to be built around — and never asked whether a stage could keep the real corpus and empty only the thing under test.

The stage

REAL_CORPUS: cwd=ROOT so t27c fpga-build --smoke resolves the real specs and succeeds, T27_ELAB_ROOT pointed at a planted tree holding a built compiler, an empty generated/ and no baseline. The gate gets past the smoke build and then finds nothing to compare against.

Reached in under a second. The real tree is only read — git status clean before and after, asserted while developing.

What came with it

  • two negative controls for the new runner, VACUOUS and WRONG, mirroring the two that certify run_on_empty_tree
  • a real 1.2 MB corpus planted in both end-to-end trees, so the new row measures wiring rather than the absence of specs
  • an UNRUN verdict when no corpus is present, matching this file's existing doctrine: a row that cannot run proved nothing, and calling it a pass is the vacuous green the file exists to catch
  • UNCOVERED 1 → 0, kept as a constant rather than deleted, because the count is asserted in the OK line and a reader who watches it fall learns something a removed line cannot tell them

Result, and its boundary

90 mutants across 13 gates — 36 silent, 21 loud, 33 invert — every one killed by the gate's own control.

What that does not say: that the gates are correct, or that their checks are the right checks. It says no mutant in these three families survives, and the families are narrow by construction. A fourth operator would be a fourth question, and the honest prior after this campaign is that a new question finds something — twice now it found the instrument.

Refs #2492

The last surviving mutant in the suite was check_elab_ratchet.py's
"no baseline" branch, declared UNCOVERED with this reason:

  "Reaching it needs `t27c fpga-build --smoke` to SUCCEED and then find
   no baseline, and a smoke build that succeeds needs the real spec tree
   -- not something an empty directory can be given."

Every clause is true. The conclusion does not follow. A control does not
have to use an empty directory: the note reasoned inside the frame of
run_on_empty_tree(), the helper this file is built around, and never
asked whether a stage could keep the real corpus and empty only the
thing under test.

REAL_CORPUS does that -- cwd=ROOT so the smoke build finds real specs
and succeeds, T27_ELAB_ROOT a planted tree with a built t27c, an empty
generated/ and no baseline. The branch is reached in under a second and
the real tree is only read (git status clean before and after).

UNCOVERED 1 -> 0. Two negative controls for the new runner (VACUOUS and
WRONG), plus a real 1.2 MB corpus planted in both end-to-end trees, plus
an UNRUN verdict when no corpus is present -- a stage that cannot run
proved nothing.

Refs #2492
@github-actions

Copy link
Copy Markdown
Contributor

PR Dashboard

Generated at: 2026-08-23 05:16:31 UTC

Summary

Status Count
Total Open PRs 6
PRs with Failing Checks 2
PRs with All Checks Green 4
READY 3
FAILING 2
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).

@github-actions

Copy link
Copy Markdown
Contributor

📓 NotebookLM Notebook linked to this PR

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

@gHashTag
gHashTag merged commit 385abf8 into master Aug 23, 2026
29 of 31 checks passed
@gHashTag
gHashTag deleted the fix/last-survivor branch August 23, 2026 05:18
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