Separate profiling failure modes and persist full audit metadata

[lullu57]

Summary

The profiling adapter collapses distinct failure modes into unsupported_counter and can misclassify truly unsupported metrics as profile_failed when ncu exits non-zero. It also omits important metadata such as profiler version and evidence-tier semantics from persisted profile records.

Evidence

• src/kernel_tuner/profiling/adapter.py:96
• src/kernel_tuner/profiling/adapter.py:204
• src/kernel_tuner/profiling/adapter.py:225
• docs/research/04_signal_and_profiling_plan.md:102
• docs/specs/profiling_adapter.md:68
• docs/specs/profiling_adapter.md:91

Why This Matters

The research plan requires a clear distinction between unsupported counters and counters missing for other reasons. Missing metadata also makes later auditability weaker.

Expected Fix

• Separate parser failure, regex miss, unsupported metric, and general profiler invocation failure.
• Persist profiler version and relevant counter-set semantics in the measurement metadata.
• Add targeted tests for failure classification.

Acceptance Criteria

• Unsupported counters are distinguishable from missing/uncollected counters.
• Persisted records include enough metadata to audit reportable vs diagnostic evidence.
• Tests cover non-zero ncu exits and missing-row cases.