scan: corpus drift against reconverge 0.3.0

[github-actions]

A scheduled scan found a recorded observation that no longer holds.

• analyzer under test: crates.io 0.3.0
• run: https://github.com/vyncint/simt-diff/actions/runs/32550692007

ANALYZER DRIFT means the same program now gets a different answer, which is
what the corpus exists to catch. GENERATOR DRIFT would mean this repository
changed instead — the generator gate ran first and passed, so that is unlikely here.

  ci-gate-lost-to-a-barrier-above              ANALYZER DRIFT
  correct-partial-warp-mask-not-gated          ok
  gate-baseline-confirmable-barrier            ok
  guard-inside-loop-not-promoted               ANALYZER DRIFT
  lane-env-barrier-below-is-harmless           ANALYZER DRIFT
  loop-inside-guard-is-promoted                ok
  truncating-cast-is-evaluated                 ok
  unreachable-barrier-still-reported           ok
  witness-reaches-only-a-prefix                ANALYZER DRIFT

entry                                        state            detail
--------------------------------------------------------------------------------------------------------------
ci-gate-lost-to-a-barrier-above              ANALYZER DRIFT   expected rc001/warning,rc001/warning|0w, got rc001/confirmed,rc001/warning|1w
correct-partial-warp-mask-not-gated          ok               signature rc002/warning|0w
gate-baseline-confirmable-barrier            ok               signature rc001/confirmed|1w
guard-inside-loop-not-promoted               ANALYZER DRIFT   expected rc001/warning|0w, got rc001/confirmed|1w
lane-env-barrier-below-is-harmless           ANALYZER DRIFT   expected rc001/confirmed,rc001/warning|1w, got rc001/confirmed,rc001/confirmed|2w
loop-inside-guard-is-promoted                ok               signature rc001/confirmed|1w
truncating-cast-is-evaluated                 ok               signature rc001/confirmed|1w
unreachable-barrier-still-reported           ok               signature rc001/warning|0w
witness-reaches-only-a-prefix                ANALYZER DRIFT   expected rc001/confirmed,rc001/warning,rc001/warning|1w, got rc001/confirmed,rc001/confirmed,rc001/warning|2w

9 entry(ies): 5 unchanged, 4 moved

ANALYZER DRIFT  ci-gate-lost-to-a-barrier-above
    finding: A warp_id()-guarded barrier ABOVE a confirmable one takes it out of the CI gate
    measured 2026-08-18 with analyzer 0.1.6
    expected rc001/warning,rc001/warning|0w, got rc001/confirmed,rc001/warning|1w
    note: Undocumented. The confirmable barrier is byte-identical to the baseline entry; only the added barrier above it differs. If this becomes confirmed, reconverge fixed it.

ANALYZER DRIFT  guard-inside-loop-not-promoted
    finding: A divergent guard inside a loop is seen but never witness-promoted
    measured 2026-08-18 with analyzer 0.1.6
    expected rc001/warning|0w, got rc001/confirmed|1w
    note: Undocumented: reconverge's Limitations section does not mention loops. Pairs with loop-inside-guard-is-promoted.

ANALYZER DRIFT  lane-env-barrier-below-is-harmless
    finding: The same barrier added BELOW changes nothing: the effect is ordered, not function-wide
    measured 2026-08-18 with analyzer 0.1.6
    expected rc001/confirmed,rc001/warning|1w, got rc001/confirmed,rc001/confirmed|2w
    note: This case refuted the first version of the ordering rule. It exists to keep the claim narrow.

ANALYZER DRIFT  witness-reaches-only-a-prefix
    finding: A, B, A: the third site shares the witnessed source and is still not promoted
    measured 2026-08-18 with analyzer 0.1.6
    expected rc001/confirmed,rc001/warning,rc001/warning|1w, got rc001/confirmed,rc001/confirmed,rc001/warning|2w
    note: The case that refuted the third and fourth versions of the ordering rule. Minimized to three one-line guards; see cases-exp-a/*/minimized/MINIMIZED.md.

Next: decide per entry whether the new behaviour is an improvement (re-record
the entry and say why in the PR) or a regression (file it upstream with the
packaged reproducer).

[vyncint]

Triaged. This is an improvement, not a regression, and it is one behaviour change rather than three — #7 and #8 carry the identical drift set and have been closed in favour of this.

Read against the oracles, which is the only way a drift report means anything:

entry	oracle	moved
ci-gate-lost-to-a-barrier-above	KNOWN_UNSAFE	warning → confirmed, +1 witness
guard-inside-loop-not-promoted	KNOWN_UNSAFE	warning → confirmed, +1 witness
lane-env-barrier-below-is-harmless	KNOWN_UNSAFE	warning → confirmed, +1 witness
witness-reaches-only-a-prefix	KNOWN_UNSAFE	warning → confirmed, +1 witness
unreachable-barrier-still-reported	KNOWN_SAFE	unchanged
correct-partial-warp-mask-not-gated	KNOWN_MASK_VALID	unchanged

Everything that moved is KNOWN_UNSAFE moving toward confirmed — the analyzer proving what it previously only suspected. Nothing moved in the column where movement would have been a false positive. That is the shape of a precision gain with no soundness cost in the direction that matters.

ci-gate-lost-to-a-barrier-above's own note anticipated it: "If this becomes confirmed, reconverge fixed it." It did.

Work split out:

• corpus: re-record the four drifted entries against reconverge 0.3.0 #11 — re-record the four entries against 0.3.0, with notes that name the version that closed each finding. Two of the entries are notes about limitations reconverge's own docs may still claim, which is an upstream documentation fix nobody knows to make yet.
• corpus: lane-env-barrier-below-is-harmless no longer shows what it was recorded to show #12 — lane-env-barrier-below-is-harmless needs more than a re-record: its finding was a distinction between two barrier positions, and both are now confirmed, so the signature can no longer show it either way.

Staying open until #11 lands, since this is the report for the version under test.