chore(pins): move the conformance pin to cuda-oxide b0f961df

[vyncint]

Closes #140. Twenty-one commits, same nightly — so this is the pin alone and not a toolchain move. Its own commit, per docs/RELEASING.md, with every gate re-run at the new rev rather than assumed.

The baseline diff: seven new gating findings, nothing lost

> conformance_unroll_bounds_check	RC003	control	deny
> conformance_unroll_bounds_check	RC003	full_after	deny
> conformance_unroll_bounds_check	RC003	full_before	deny
> conformance_unroll_bounds_check	RC003	full_inside	deny
> conformance_unroll_bounds_check	RC003	partial_after	deny
> conformance_unroll_bounds_check	RC003	partial_before	deny
> conformance_unroll_bounds_check	RC003	partial_inside	deny

No < lines: nothing that was detected stopped being detected.

conformance_unroll_bounds_check is upstream's new regression corpus for a bounds-check bug in the #[unroll] MIR transform. All seven kernels take arr: &mut [u32] precisely so the slice can be indexed out of range (arr[999] = 7) — the point of the example is the missing check, not the parameter. RC003 is correct on every one of them: &mut [T] is one exclusive reference handed to every thread, which is also what upstream's own safety model says. Reviewed true positives, recorded in EXPECTED with that reason, next to the existing constant_index_from_end entry which is the same shape.

The surface gate: one new primitive, classified

unknown: cuda_device::thread::__grid_constant_config
check-surface: FAIL — 1 upstream surface function(s) neither classified nor allowlisted

It is a #[doc(hidden)] compile-time launch-ABI marker for #[grid_constant], #[inline(never)] with an empty body — "detected at compile time and removed, no runtime code is generated". It emits no instruction and has no participation semantics, so it is allowlisted with that reason rather than left as Other by default, which is what the policy asks.

The mutation corpus: regenerated, and the numbers improve

class	before	after (default)	before	after (--strict)
wrapbar	56/78 (71%)	60/78 (76%)	73/78 (93%)	77/78 (98%)
wrapcol	23/42 (54%)	23/35 (65%)	42/42 (100%)	35/35 (100%)
mutslice	424/424 (100%)	434/434 (100%)	—	—
shrinkmask	0/17 (0%)	0/10 (0%)	—	—

Precision stays 1.000, now over 576 gating findings rather than 506.

This is a different population, not the same mutants scoring better: upstream rewrote examples/atomics to take *mut u32 with DeviceAtomicU32::from_ptr instead of transmuting a shared slice, and edited coop_groups_demo — so the kernels the corpus is generated from changed. The table is regenerated by the script and committed, never hand-edited.

That rewrite is also why the informational line reads 4 RC001/RC002 chains complete rather than 8: the upstream example this project had recorded as a likely-real finding (atomics::atomic_i32_test, in docs/hardware/session-1.md) no longer has the shape that produced it. Every remaining chain is still complete, which is the half that actually gates.

Verified locally

conformance: PASS — gating findings match the baseline exactly
check-surface: PASS — every surface function is classified or allowlisted
check-surface self-test: PASS — the gate fails on every input that should fail
mutation: PASS — precision 1.0 at default confidence; published table unchanged

plus cargo test --workspace with no failures. The three sample-crate manifests and their lockfiles move with the pin; render-probe gains the Cargo.lock it had never committed.