Skip to content

Two of the nine recorded survivors are killable; a third cannot be killed at all - #47

Closed
d6g8k5htny-coder wants to merge 1 commit into
claude/cover-uniform-cost-contractfrom
claude/cover-survivor-tangency
Closed

d6g8k5htny-coder wants to merge 1 commit into
claude/cover-uniform-cost-contractfrom
claude/cover-survivor-tangency

Conversation

@d6g8k5htny-coder

Copy link
Copy Markdown
Owner

Stacked on #46 (→ #38 → migration). Two files: three controls and a README correction.

The defect

research/cover/README.md listed nine surviving mutants under one dismissal — six "exact-equality boundaries that the reference geometry never lands on" — and closed by promising that "a cover whose cell widths happened to hit one of those equalities would separate them."

One table row bundled two different sites under a single description:

regions.py:399, :412 (x2) | <= → <, < → <= | disjointness tests, differing only at exact tangency.

That description is wrong for one site and too weak for the other, and the two errors point in opposite directions.

:412 — killable, and the equality is reachable

if hi2 < rlo2 or lo2 > rhi2:
    return OUTSIDE

A genuine disjointness test. Exact rationals make a witness trivial — [0, 3/50] × [0, 4/50] straddles the origin on both axes, so hi2 = 9/2500 + 16/2500 = 1/100 = r_lo² exactly:

box live mutant < → <= mutant > → >=
inner-tangent, hi2 = r_lo² STRADDLE OUTSIDE —
outer-tangent, lo2 = r_hi² STRADDLE — OUTSIDE

A cell touching the annulus, rejected as proved disjoint, whose residual total() then discards outright (see control 41 on #38). Each box kills exactly one comparison, so both controls are needed. Controls 44 and 45, each confirmed to FAIL against a scratchpad copy with the comparison flipped.

:399 — not a disjointness test, and killable by nothing

lo = Fraction(0) if (a <= 0 <= b) else min(abs(a), abs(b))

The per-axis min/max helper. Its two operand flips differ nowhere on a legal box, not "only at tangency":

  • a < 0 <= b can diverge only when a == 0 — the else branch then returns min(|0|, |b|) = 0, which is what the if branch returns.
  • a <= 0 < b likewise when b == 0 — min(|a|, 0) = 0.

Verified over 797 legal pairs, zero divergences.

So the README's closing promise is now demonstrated for :412 and false for :399, where no cover, no cell width and no input can separate them. Control 46 asserts the equivalence rather than pretending a kill is available — for an equivalence claim that is the only guard there is. If _axis_min_max ever changes so the flips disagree, the README's claim becomes false and the control fails.

What I did not do

The other six entries stay as the sweep's own classification, and the README now says so explicitly. A correction that found the sweep wrong once in each direction is not grounds for silently trusting the rest of it — but it is also not grounds for me to claim I have re-examined sites I have not. They remain open.

Verification

10/10 checkers rc=0
python3 -m pytest -q     1,931 passed, 1 skipped (574s)
tests/test_cover.py      54 passed  (48 → 50 on #38 → 51 on #46 → 54 here)

Mutation evidence, against a copy under the scratchpad and never the library: flipping both comparisons at :412 fails controls 44 and 45 while 46 correctly still passes, since it concerns a different line and asserts equivalence rather than a kill.

What this does not establish

No mathematics. Killing a mutant is a statement about a test suite, and proving one equivalent is a statement about a helper function. Every integrand here remains a labelled reference, no cell of the program's actual cover is certified, and D3-LEMMA-RN-UNIF Piece 1 and Piece 2 remain OPEN.

Third of fourteen survey survivors; nine remain.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY


Generated by Claude Code

…lled at all

`research/cover/README.md` listed nine surviving mutants under one dismissal:
six "exact-equality boundaries that the reference geometry never lands on", and
a closing sentence promising that "a cover whose cell widths happened to hit one
of those equalities would separate them". One table row bundled
`regions.py:399` and `:412` under a single description -- "disjointness tests,
differing only at exact tangency" -- and that description is wrong for one site
and too weak for the other. The two errors point in opposite directions.

`:412` IS a disjointness test, `if hi2 < rlo2 or lo2 > rhi2: return OUTSIDE`,
and the equality is reachable. Exact rationals make a witness trivial:
[0, 3/50] x [0, 4/50] straddles the origin on both axes, so hi2 =
9/2500 + 16/2500 = 1/100 = r_lo^2 exactly. Live it is STRADDLE; mutated it is
OUTSIDE -- a cell touching the annulus, rejected as *proved disjoint*, whose
residual total() then discards outright. The mirror box [5, 6] x [0, 1] gives
lo2 = 25 = r_hi^2 and kills the other comparison. Each box kills exactly one,
so both controls are needed. Controls 44 and 45, each confirmed to FAIL against
a scratchpad copy with the comparison flipped.

`:399` is not a disjointness test at all. It is the per-axis min/max helper,
`lo = Fraction(0) if (a <= 0 <= b) else min(abs(a), abs(b))`, and its two
operand flips differ NOWHERE on a legal box rather than only at tangency:
`a < 0 <= b` can diverge only when a == 0, and the else branch then returns
min(|0|, |b|) = 0, which is what the if branch returns; `a <= 0 < b` likewise
when b == 0. Verified over 797 legal pairs, zero divergences.

That distinction is the point rather than a pedantic one. The README's closing
promise -- that a cover landing on the equality would separate them -- is now
DEMONSTRATED for :412 and FALSE for :399, where nothing can separate them.
Control 46 therefore asserts the equivalence instead of pretending a kill is
available, which is the only guard an equivalence claim can carry: if
`_axis_min_max` is ever changed so the flips disagree, the README's claim
becomes false and the control fails.

The README row is split, each site described accurately, and the closing
paragraph rewritten to say what was corrected and in which direction. The other
six entries are left as the sweep's own classification, explicitly not
re-examined here -- a correction that found the sweep wrong once in each
direction is not grounds for silently trusting the rest of it.

This establishes no mathematics. Killing a mutant is a statement about a test
suite; D3-LEMMA-RN-UNIF Piece 1 and Piece 2 remain OPEN.

10/10 checkers pass; 1,931 passed, 1 skipped.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
d6g8k5htny-coder pushed a commit that referenced this pull request Oct 20, 2026
…uery-#47

The Ready review of query-#47 at 4fb15c97 found three portability
problems in the class this session wrote. This adds an author-side
correction under mutation/query_archive_v2/:

- force-add the descendant fixture paths, which an excludes file may
  ignore;
- skip the SHA-256 test when Git cannot create a SHA-256 repository;
- bound repository discovery with GIT_CEILING_DIRECTORIES in the
  non-repository test.

Also included: the patch against #47's head, a harness that reproduces
each finding, and its local outputs for both class versions together with
the mutant control (10/10 killed in both modes). README, MANIFEST.json
and SHA256SUMS are updated. Retention only, not for merge. Scientific
effect NONE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013tN78b3Q38WdttFyiJ6KVg
d6g8k5htny-coder pushed a commit that referenced this pull request Oct 20, 2026
An OpenAI Codex writer fixed the same three portability findings on
query-#47 at a378fa41. That head passes all four configurations of the
retained repro.py in normal and -O mode, so this session's correction
was not posted or applied. README and MANIFEST.json now say so, and
SHA256SUMS is regenerated. Retention only. Scientific effect NONE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013tN78b3Q38WdttFyiJ6KVg
d6g8k5htny-coder added a commit that referenced this pull request Nov 9, 2026
Remove the unexercised historical acceptance driver from the operational tool surface, keep its exact Git history, and localize the byte-identity helpers in the two current qualified browser checks. Bind the five canonical CBZ surfaces to Main #773, Workspace #47, durable v135, and the exact 23-tool G11 residue without changing proof or scientific status.
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