Add lazy loading support for code objects

[mawad-amd]

Summary

• Add addFile(path, lazy=true) mode that indexes kernel names from ELF symbols without disassembling.
• Add ensureKernelLoaded(name) to disassemble a single kernel on demand when it is first dispatched.
• Enables downstream tools (e.g. Nexus) to defer the expensive disassembly step until a kernel is actually needed.

🤖 Generated with Claude Code

[rwvo]

This review was written by Claude on behalf of Rene van Oostrum.

Build verification

Omniprobe builds cleanly against this branch and all 27 tests pass (22 handler + 5 Triton integration). The lazy=false default preserves existing behavior — API-compatible, ABI-breaking (acceptable for a submodule built from source).

Lazy loading is interesting for Omniprobe

Omniprobe currently calls scanCodeObject() on first kernel dispatch, which disassembles the entire code object. For hipBLASLt's matrix transform .hsaco (960 kernels), this is expensive when only a handful of kernels are actually dispatched. Switching to addFile(path, agent, filter, lazy=true) would let Omniprobe defer disassembly to per-kernel on-demand loading — a natural fit for its dispatch-driven architecture.

That said, the following issues in the lazy path would need to be resolved first.

Must fix

1. Thread safety in ensureKernelLoaded() — Two threads calling getKernel("foo") concurrently will both find foo in lazy_kernels_ (shared lock), both disassemble it (no lock), and both call addKernel(). The second addKernel() replaces the unique_ptr, potentially destroying the CDNAKernel that the first thread already holds a reference to via the returned CDNAKernel&. kernelDB already exposes a shared_mutex contract for thread safety — the lazy path needs to uphold it (e.g., a "loading in progress" sentinel so the second thread waits).

2. Temp file leak in getDisassemblyForSymbol() — If invokeProgram() returns false, the temp file created by tmpnam() is never unlinked.

3. Unconditional std::cerr debug output — Five [KDB] prints to stderr on every lazy load. Should be behind a debug flag or removed.

Should fix

4. No test coverage for the lazy path — At minimum, a Python test that does add_file(hsaco, lazy=True), verifies has_kernel() returns True, then calls get_kernel() and checks the result matches eager-loaded output.

5. scanCodeObjectForKernel / parseDisassemblyForKernel should be private: — Only called from ensureKernelLoaded().

6. Silent failure in ensureKernelLoaded() — If disassembly fails, the kernel stays in lazy_kernels_ forever. Every subsequent getKernel() call re-attempts the disassembly, and the caller only sees a misleading "kernel does not exist" exception with no indication the real cause was a disassembly failure. Either remove the entry on failure (fail fast) or propagate the error.

7. Replace tmpnam() with mkstemp() — tmpnam is TOCTOU-racy; the rest of the codebase uses mkstemp-style patterns.

8. Code duplication — parseDisassemblyForKernel duplicates ~170 lines from parseDisassembly. Consider adding an optional targetKernel parameter to the existing function (as was already done for processKernelsWithAddressMap).

Nice to have

9. getDisassemblyForSymbol is missing __attribute__((visibility("default"))) — unlike getDisassembly which has it. If internal-only, remove from the public header.
10. The pyproject.toml license format change is unrelated — consider splitting into its own PR.
11. Bare catch(...) blocks swallow all exceptions silently. Catching specific types (std::invalid_argument, std::out_of_range) and logging the malformed input would help debugging.

[mawad-amd]

Thanks Rene for the thorough review! My agent thanks your agent for the thorough review. 🤖🤝🤖 (We're one review cycle away from the agents just cutting us out entirely.)

All must-fix and should-fix items have been addressed, CI is green ✅. Please give it a spin with Omniprobe before we merge to make sure everything's working as intended. Thanks!

[rwvo]

My agent was happy with the work of yours and merged the PR.