Summary

hipBLASLt and hipSPARSELt each compiled their own copy of the TensileLite host code, so
both libhipblaslt.so and libhipsparselt.so exported the same ~117 TensileLite symbols.
Under ELF flat-namespace interposition the loader binds all references to whichever copy
loads first — the cause of a March 2026 segfault (#6917, originally worked around by hiding
all TensileLite symbols at the source level with bespoke TENSILE_HIDDEN_BEGIN/END
wrappers).

This PR makes tensilelite-host a first-class, exportable CMake target with a curated
public API, and removes the source-level symbol-hiding workaround in favor of the standard
generated export header. It also folds in the device-library codegen export and the
hipSPARSELt consumption path, so the whole "TensileLite host is a real library, not an
internal blob" change lands as one unit.

Default build mode is static embed. tensilelite-host builds STATIC by default and is
absorbed into each consumer; under the hidden-visibility preset its internals (and, in
static mode, the public API itself) stay hidden, so neither consumer re-exports the
TensileLite surface and the interposition hazard is avoided — the same end state as the old
workaround, achieved through the normal export-header mechanism rather than hand-maintained
hidden regions. A single-owner shared libtensilelite-host.so (true structural dedup) is
available behind TENSILELITE_BUILD_SHARED_LIBS=ON but is not the default yet: it needs
a single producer across the two consumers plus a TheRock packaging edge that is still being
worked out, so it is deferred to avoid shipping a half-wired shared mode.

What changed

Library shape & exports

• tensilelite-host is an exportable target (was an OBJECT library absorbed into each
  consumer). Build type is driven by TENSILELITE_BUILD_SHARED_LIBS (default OFF → static
  embed; ON → co-exported roc::tensilelite-host shared library). The dedicated option is
  used instead of the not-yet-synced global BUILD_SHARED_LIBS so the fork can't silently
  flip the mode.
• Export inversion, not a rename. TENSILE_API flips from a hidden marker to the
  generated TENSILELITEHOST_EXPORT (generate_export_header). Under the hidden visibility
  preset only the curated public API is annotated for export; internals stay hidden. In a
  static build the macro resolves empty (TENSILELITEHOST_STATIC), so the public API stays
  hidden inside the consuming .so too. The 100 dead TENSILE_HIDDEN_BEGIN/END no-op
  wrappers are removed.
• Stop re-exporting static LLVM. With YAML serialization on, tensilelite-host links the
  static LLVM archives; as a shared library it re-exported ~3700 LLVM symbols at default
  visibility (the preset governs the target's own code, not prebuilt archives). In a client
  process those interpose on the HIP runtime's in-process comgr LLVM and corrupt its AMDGPU
  cl::opt registry — comgr then rejects its own backend options at code-object load
  (Unknown command line argument '-amdgpu-prelink'). Linking with --exclude-libs,ALL
  (scoped to SHARED AND NOT WIN32) localizes the static-archive symbols; LLVM exports drop
  ~3700 → ~10 with the public API unchanged.
• Windows dllexport fix. Annotate the friend operator<<(TensorDescriptor) first
  declaration with the export macro so clang-cl doesn't error on a redeclaration.

Device-library codegen (formerly "Phase 3")

• Codegen is an exported CMake function. New HipBLASLtCodegen.cmake /
  HipBLASLtCodegenInstall.cmake modules wire the Tensile device-library generation as a
  reusable, exported step so a downstream consumer can produce TensileLibrary_lazy_<arch>
  without re-deriving the codegen invocation.
• hipSPARSELt consumes hipBLASLt via find_package on TheRock. When hipBLASLt is a
  provided package (HIPSPARSELT_ENABLE_THEROCK AND "hipblaslt" IN_LIST THEROCK_PROVIDED_PACKAGES) hipSPARSELt does find_package(hipblaslt CONFIG REQUIRED);
  otherwise it builds hipBLASLt in-tree via add_subdirectory.

Install-surface correctness

• Ship the Tensile device library; keep embedded deps out of the package. EXCLUDE_FROM_ALL
  on the in-tree hipBLASLt subdirectory was suppressing the add_custom_target(... ALL) that
  generates the device library, so TensileLibrary_lazy_<arch>.dat never installed and the
  hipSPARSELt suite failed every case with "Could not initialize Tensile library". Root fix:
  gate origami's and stinkytofu's install/export/package rules on <PROJ>_STANDALONE
  (top-level project ⇒ find_package'd ⇒ installs its surface; add_subdirectory ⇒ embedded
  ⇒ installs nothing), matching what rocisa already does. hipSPARSELt then drops
  EXCLUDE_FROM_ALL: hipBLASLt's device-library target builds and installs to
  lib/hipsparselt/library with no origami/stinkytofu pollution in the install tree.
• Headers stay in-tree for the library export. tensilelite-host ships binary-only; the
  Tensile host header file sets are attached for build-tree/IDE organization and consumed via
  BUILD_INTERFACE, so the installed Tensile/*.hpp collision problem cannot occur.
• Removed the dead TENSILE_USE_HIP non-HIP DataTypes paths, un-gated
  TENSILE_DEFAULT_SERIALIZATION, deleted the dead Tensile-fork glue and the old hipSPARSELt
  per-symbol workaround, and kept origami out of the installed package surface.

Diagnostics

• Structured diagnostics facility. A [tensilelite:diag] logfmt+banner facility in the
  client (Diagnostic.hpp) mirrored in the python test harness, with auto config/gpu/phase
  context, to make client/codegen failures legible in CI.

CI fixes folded in

• gfx950 multi-DU YAMLs: drop dead GlobalParameters keys that fail validation on the
  current loader.
• joblib Set changed size during iteration: the logic-file load used
  ParallelMap2(return_as="generator_unordered"). joblib ≥1.5.0's unordered _retrieve
  iterates self._jobs_set unlocked (next(iter(...))) while the dispatcher thread mutates
  it, racing into RuntimeError: Set changed size during iteration (seen on the gfx950 /
  ubuntu-24 precheckin loading 648 logic files across 32 threads). Switched to ordered
  return_as="generator", which waits on self._jobs[0] and never iterates _jobs_set.
  Result is unchanged — the loop merges into masterLibraries and sorts explicitly for
  determinism, so consumption order is irrelevant. (Latent since the call adopted unordered
  mode in Jan 2025 [hipBLASLt] Remove debug parameter #1514; armed once CI picked up joblib ≥1.5.0.)

Validation (gfx90a / gfx942)

• Static default: tensilelite-host builds libtensilelite-host.a; the export carries it
  STATIC IMPORTED with INTERFACE_COMPILE_DEFINITIONS TENSILELITEHOST_STATIC. The public
  API stays hidden inside the consuming .so, so neither consumer re-exports the TensileLite
  surface.
• Shared opt-in (TENSILELITE_BUILD_SHARED_LIBS=ON): libhipblaslt.so DT_NEEDEDs the
  single libtensilelite-host.so.1; 0 strong-symbol duplicates between them. Remaining
  overlaps are benign weak template/inline vague-linkage, loader-deduped. hipSPARSELt
  resolves the same single lib.
• Symbol surface: static LLVM is localized — relink shows ~3700 → ~10 LLVM symbols
  exported, 0 cl::opt/target-registry symbols remain, TensileLite API exports unchanged.
• Device library installs through the add_subdirectory path (gfx942):
  TensileLibrary_lazy_gfx942.dat.zlib installs to lib/hipsparselt/library; the install
  tree contains no origami/stinkytofu headers, cmake exports, or libraries.

TheRock-visible delta

• No new top-level package. With the shared mode enabled, roc::tensilelite-host is
  co-exported inside hipBLASLt's existing package; in the default static build there is no new
  installed library at all.
• No new installed headers — Tensile host headers stay in-tree.
• New hipSPARSELt → hipBLASLt edge: hipSPARSELt either find_packages hipBLASLt (TheRock,
  provided) or builds it in-tree, and the device library now installs to lib/hipsparselt.
• Origami/stinkytofu install their package surface only when built standalone; embedded into a
  consumer they install nothing.

AIHPBLAS-3522