GHC 9.12.2: UTxO read regression (2x on 10k) and heavier TxHistory writes

[paolino]

Context

Benchmark comparison across GHC 9.6.6, 9.8.2, 9.10.1, 9.12.2 (see #5166) revealed two regressions in 9.12.2.

UTxO Read — 2x regression on 10k entries

Workload	9.6.6	9.8.2	9.10.1	9.12.2
10k ada-only	0.297ms	0.312ms	0.335ms	0.645ms
100k ada-only	2.74ms	2.94ms	3.79ms	3.36ms

The 10k case doubles from 0.297ms to 0.645ms. The 100k case shows a smaller ~23% regression.

TxHistory Write — ~11% regression on heavy workloads

Workload	9.6.6	9.8.2	9.10.1	9.12.2
1000 50i+100o (run 3)	61.6s	59.4s	63.4s	66.1s
10000 10i+10o	6.85s	6.70s	7.19s	7.47s

Notes

• Both are DB-bound workloads where serialization/deserialization code may be affected by GHC codegen changes
• 9.12.2 is the overall winner on every other benchmark (memory -7.5%, read-blocks, API, latency)
• 9.8.2 was consistently best on these DB write benchmarks across all GHC versions tested
• Worth investigating whether this is a GHC codegen issue or a CBOR/store serialization path that changed behavior

Benchmark runs

• 9.6.6
• 9.8.2
• 9.10.1
• 9.12.2

Acceptance criteria

• Root cause identified and documented (linked PR or comment) for the 10k UTxO-read 2x regression on 9.12.2
• UTxO Read 10k ada-only on GHC 9.12.2 within 20% of the GHC 9.6.6 baseline (0.297ms) in benchmark CI
• TxHistory Write 1000 50i+100o on GHC 9.12.2 within 5% of the GHC 9.8.2 baseline (59.4s)
• Follow-up benchmark run linked in this issue demonstrating the fix on main