feat(accel): exclusive capabilities, an open accel grammar, and the deprecated environment key removed from the manual

[Sunrisepeak]

Three things from the multi-device design
(.agents/docs/2026-09-05-multi-device-ecosystem-design.md).

1. [package] exclusive — closing the same-symbol case

Two packages providing one capability is usually fine and sometimes the point:
an OpenBLAS package and an MKL package both provide blas, and a build that
links one program against each is legitimate. The manual says so.

What is not fine is two implementations of one accelerator interface in one
link — they define the same symbols, and the link resolves every call to
whichever archive it reached first. The manual documents that failure too, with
a measured multiple definition of 'cap_probe'.

The engine cannot tell those apart: seeing the symbol overlap needs object files
that do not exist when capabilities are bound. So the package says it.

provides  = ["gpu-blas"]
exclusive = ["gpu-blas"]

error: capability 'gpu-blas' is provided by more than one package, and they
       declare it EXCLUSIVE.
         providers: [gpublas-a, gpublas-b]
         exclusive: [gpublas-a, gpublas-b]
       Two implementations of one interface define the same symbols, so the
       link would resolve every call to whichever archive it reached first.
       Keep one of them — a `[capabilities]` pin selects a provider for a
       REQUIREMENT and cannot make two definitions of one symbol safe.

Reported as exclusive-capability in --format json. A capability nobody
declares exclusive behaves exactly as before.

A list rather than a flag, because a package may provide several capabilities
and be the sole provider of only some. An entry naming a capability the package
does not supply is a schema warning — exclusivity is a claim about this
package's own symbols, and a silent no-op is how a typo survives.

2. The accel grammar is open

The backend set was never closed, but the floor had only CUDA's spelling.
floor>= is now accepted alongside ptx>= for the same field.

vulkan1.3+{spirv1.6} floor>=1.4
sycl2020+{spir64,nvptx64-sm_89}
hip6.4+{gfx942}

An architecture carrying no leading number — gfx942 — is compared by
equality, because there is no ordering to read out of it. Reading 942 out of
the middle would invent a level AMD does not define.

3. [xlings] deps removed from the manual

[xlings.workspace] has been the one table since 2026.9.3.1, and deps is
honoured with a warning that names the replacement. The manual still taught the
old spelling in fourteen places in English and ten in Chinese, and
examples/09-cuda-kernel — shipped in 2026.9.5.1 — was written with it.

A deprecated form that still works is exactly the one a reader copies, so the
documentation no longer offers it. The parser's compatibility path is
untouched.

Verified

tests/e2e/601_exclusive_capability.sh asserts both halves, and the second
is the one that matters:

two providers, both declaring exclusive	refused, naming the capability and both
the same graph with no claim	still builds

Without that control the test would also pass against an engine that refused
every duplicate provider — the behaviour the key exists to avoid.

100 test binaries pass. The three new manifest tests and five new grammar tests
were run individually to confirm they execute rather than merely being counted:

[  PASSED  ] 3 tests.     ManifestExclusive.*
[  PASSED  ] 5 tests.     AccelOpenGrammar.*

Companion

Payloads land separately in openxlings/xim-pkgindex#759 (25 packages: the CUDA
toolkit split by upstream component, plus DPC++). They do not depend on this
change and this change does not depend on them.