perf: revert the tuning experiment and record the negative result

The D-047 amortization change is reverted: it is measurably correct and
does what it was designed to do, but it does not earn its place.

Component microbenchmarks (5 repetitions, medians) showed the predicted
mechanism clearly: FastBook add at an existing level 42.37 to 27.03 ns,
add creating a level 50.84 to 43.09 ns, delete 18.07 to 14.57 ns.
Paired alternating full-file replays told a different story at scale:
about 4.5 percent faster at 50 million frames, but medians within about
1 percent with overlapping distributions at 100 million frames, the
scope closest to the real working set. Peak anonymous footprint rose
about 14 percent, from 243 to 276 MiB over the same scope, exactly the
regression the hypothesis predicted. Sparser tables span more cache
lines, so the locality cost appears to cancel the rehash saving while
the memory cost stays.

The acceptance rule written before the experiment required improvement
beyond run-to-run noise with no memory regression, and neither holds, so
the change is reverted and D-048 records the full evidence. Correctness
was green throughout, including the 100-million-message dual-engine
validation. The tuned full-protocol run is preserved as
bench/results/2026-07-26-tuned-experiment.json; it cannot decide the
question because its own controls (the book-free framing scan and the
unchanged MapBook) moved more than the effect under test.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: IpsitKalra

bench/results/2026-07-26-tuned-experiment-env.txt

@@ -0,0 +1,52 @@
+capture_local_time: 2026-07-26 17:48:04 -0400
+capture_utc_time: 2026-07-26T21:48:04Z
+host_label: Mac
+operating_system: Darwin
+kernel_release: 25.5.0
+kernel_version: Darwin Kernel Version 25.5.0: Tue Jun  9 22:27:52 PDT 2026; root:xnu-12377.121.10~1/RELEASE_ARM64_T8112
+cpu_architecture: arm64
+os_product_version: 26.5.2
+os_build_version: 25F84
+cpu_model: Apple M2
+cpu_physical_cores: 8
+cpu_logical_cores: 8
+cpu_performance_cores: 4
+cpu_efficiency_cores: 4
+memory_installed_bytes: 8589934592
+memory_free_bytes: 58589184
+memory_inactive_bytes: 1315667968
+swap_usage: total = 9216.00M used = 8175.62M free = 1040.38M (encrypted)
+power_source: Battery Power
+battery_state: -InternalBattery-0 (id=37552227) 54%; discharging; 2:16 remaining present: true
+low_power_mode: 0
+thermal_pressure: unavailable
+thermal_warnings: none recorded
+dataset_basename: 07302019.NASDAQ_ITCH50
+dataset_size_bytes: 8661679413
+dataset_filesystem: 
+dataset_sha256: 9f8634a048b8195ccdcbe618e5833a759e35e04486d74e36b79176d35172654a
+dataset_path_recorded: no (basename and checksum only, by policy)
+git_commit: dfa1ed2acfbbb8a67f7dabc3f65c3e75d18be57f
+git_short_commit: dfa1ed2
+git_branch: main
+git_working_tree: clean
+repository: IpsitKalra/nanobook
+compiler_executable: c++
+compiler_version: Apple clang version 16.0.0 (clang-1600.0.26.6)
+build_type: Release
+release_flags_base: -O3 -DNDEBUG
+native_arch_option: ON
+lto: ON
+host_tuning_flag: -march=native
+cmake_version: 4.4.0
+cxx_standard: C++20
+cpu_pinning_available: no (taskset absent on this platform)
+cpu_pinning_used: no
+selected_cpu_core: none
+cpu_governor: unavailable on this platform
+perf_available: no
+xctrace_available: no (binary present but requires full Xcode)
+sample_utility_available: yes
+platform_class: documented macOS fallback (product_spec.md section 1)
+tsc_expected: no (no x86 TSC on this architecture)
+disk_free_bytes_repo_volume: 15108644864