test(bench-orchestrator): replace golden matrix snapshots with targeted assertions

The golden files pinned every field of all nine profiles, so any intentional
matrix change meant regenerating nine JSON snapshots -- churn with the standard
snapshot failure mode of blind regeneration. Drop them in favor of targeted
assertions on the load-bearing fields (selection and engine/format targets), a
structural smoke check that every profile resolves to well-formed entries, and a
check on the secondary fields CI depends on (remote_key, scale_factor). The
develop/nightly behavior-preservation tests are unchanged.

Signed-off-by: Connor Tsui <connor@spiraldb.com>
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PQg1cgPLKJg4QVJvE2xa3Q

Author: claude

bench-orchestrator/README.md

@@ -235,12 +235,13 @@ jobs:
             ${{ matrix.scale_factor && format('--opt scale-factor={0}', matrix.scale_factor) || '' }}
 ```
 
-### Golden and parity tests
+### Tests
 
-`tests/test_matrix.py` checks each profile against a checked-in golden file
-(`tests/golden/<profile>.json`), so any change to what-runs surfaces as a reviewable diff.
-Regenerate them with `UPDATE_GOLDEN=1 pytest`. Two further tests assert that the `develop` and
-`nightly` profiles reproduce the current workflow matrices exactly, proving the registry is
+`tests/test_matrix.py` covers each profile with targeted assertions on its load-bearing fields --
+which benchmarks are selected, their engine/format targets, and the secondary fields CI depends on
+(`remote_key`, `scale_factor`, `data_formats`) -- plus a structural smoke check that every profile
+resolves to a well-formed matrix. The `develop` and `nightly` profiles additionally assert they
+reproduce the target sets the old inline workflow matrices used, proving the registry is
 behavior-preserving.
 
 ## Example Workflows