Prune redundant tests and remove unused helpers

documents/LATENCY_WHITEPAPER.md

@@ -134,20 +134,12 @@ so stride must be a multiple of 8.
 
 The following behaviors are explicitly covered by unit tests:
 
-- **Null buffers, zero/misaligned strides, and one-node buffers are rejected with exact reasons**
-  (`SetupLatencyChainRejectsInvalidInputsWithExactReasons`)
-- **The minimum two-node chain is accepted at pointer-sized and default strides**
-  (`SetupLatencyChainAcceptsMinimumTwoNodeChainsAtSupportedStrides`)
-- **A constructed chain visits every aligned in-buffer node exactly once and returns to its head**
-  (`SetupLatencyChainCreatesValidChain`)
-- **Pointer count, touched-page count, page size, and stride diagnostics are exact**
-  (`SetupLatencyChainCollectsDiagnostics`)
-- **A locality window that cannot contain two stride-spaced nodes is rejected**
-  (`SetupLatencyChainWithTooSmallTlbLocalityFails`)
-- **A locality-using box mode with a zero locality window is rejected**
-  (`SetupLatencyChainWithBoxModeAndZeroLocalityFails`)
+- **Null buffers, zero/misaligned strides, one-node buffers, unusable locality windows, and box modes without
+  locality are rejected with exact reasons** (`SetupLatencyChainRejectsInvalidInputsWithExactReasons`).
+- **Two-node and four-node chains at pointer-sized and default strides visit every aligned in-buffer node exactly
+  once and return to the head** (`SetupLatencyChainCreatesValidChain`).
 - **Equal explicit seeds reproduce the chain, different seeds change it, and same-random box mode reuses one
-  permutation across boxes** (`ExplicitSeedsAreReproducibleAndSameRandomModeReusesBoxPermutation`)
+  permutation across boxes** (`ExplicitSeedsAreReproducibleAndSameRandomModeReusesBoxPermutation`).
 
 These tests validate setup correctness and failure handling before timing is run.
 

documents/LLM_MEMORY_PROFILE_WHITEPAPER.md

@@ -793,8 +793,7 @@ The maintained examples are in [CPU kernel tests](../tests/test_llm_memory_kerne
 [Metal helper tests](../tests/test_llm_metal_checksum.mm) for
 `SharedAffineLanesExposeBoundedContentCollisions`. The
 [Metal backend tests](../tests/test_llm_metal_backend.cpp) contain
-`DecodePagedPermutationAndPaddingHooksAreDetectedIntegration` and
-`PrefillPagedPermutationAndPaddingHooksAreDetectedIntegration`. Their real table/blit mutations differ from
+`PagedPermutationAndPaddingHooksAreDetectedIntegration`, which covers decode and prefill. Its real table/blit mutations differ from
 `force_timed_checksum_mismatch` (host readback altered after execution) and `force_kv_write_validation_mismatch`
 (forced validation boolean), which test result handling. Pure independent checksum-oracle goldens test arithmetic;
 they are not GPU execution evidence. Algorithms and their existing identities remain unchanged by this fault model.

documents/PROJECT_STRUCTURE.md

@@ -195,7 +195,7 @@ Core infrastructure for configuration, memory management, macOS system introspec
 | `buffer_manager.h` | Defines the phase-local `BenchmarkBuffers` and command-local `PatternBuffers` RAII owners |
 | `buffer_allocator.h` / `.cpp` | Atomically allocates the pattern source/destination pair and performs overflow-safe peak-concurrent memory accounting |
 | `buffer_initializer.h` / `.cpp` | Initializes deterministic pattern source data and the zeroed pattern destination before execution |
-| `memory_utils.h` / `.cpp` | Cache-line alignment helpers, deterministic/unseeded latency-chain construction and diagnostics, and basic buffer initialization |
+| `memory_utils.h` / `.cpp` | Cache-line alignment helpers, deterministic/unseeded latency-chain construction, and basic buffer initialization |
 
 #### src/core/system/
 
@@ -375,20 +375,19 @@ installed. All test `.cpp` and `.mm` files are picked up automatically by the Ma
 |---|---|---|
 | `test_config.cpp` | `ConfigTest` | Strict whole-token CLI/sweep parsing, validation, defaults, and derived buffer/access math |
 | `test_signal_handler.cpp` | `BenchmarkSignalMaskGuardTest` | Exact thread-mask restoration, caller-preserved blocked signals, and nested scope ownership |
-| `test_messages.cpp` | `Messages*Test` | Console message string functions across all categories |
+| `test_messages.cpp` | `Messages*Test` | Exact help and formatted console message contracts |
 | `test_memory_utils.cpp` | `MemoryUtilsTest` | Memory helpers: pointer-chase chain construction and verification |
 | `test_memory_manager.cpp` | `MemoryManagerTest` | Injected mmap/madvise policy, failures, and exact RAII unmapping |
 | `test_numeric_utils.cpp` | `NumericUtilsTest` | Overflow-safe arithmetic, duration calibration, pilot counts, and quantization boundaries |
-| `test_llm_memory_contract.cpp` | `LlmMemoryContractTest` | Independent executable specification for all eight decode/prefill backend/layout methodology identities, payloads, paged geometry/lookup goldens, append and full-prompt writes, checksum/event-trace identities, descriptor ABI layouts, and schema acceptance; it does not exercise production LLM code |
-| `test_llm_memory_config.cpp` | `LlmMemoryConfigTest` | Config/status defaults, exact standalone decode/prefill parsing including paged prefill, phase/layout rules and stable reasons, strict decimal errors, raw output values, help isolation, worker/seed resolution, geometry/work-limit preflight, and incompatible options |
+| `test_llm_memory_config.cpp` | `LlmMemoryConfigTest` | Exact standalone decode/prefill parsing including paged prefill, phase/layout rules and stable reasons, strict decimal errors, raw output values, help isolation, worker/seed resolution, geometry/work-limit preflight, and incompatible options |
 | `test_llm_memory_work_plan.cpp` | `LlmMemoryWorkPlanTest` | Production checked decode/prefill geometry and payloads, physical/padding/table/lookup/accounted math including paged-prefill `N+2*M`, deterministic permutation/hash, CPU token/block ownership, Metal cyclic block-owner scheduling and component identities, memory budget, descriptor/range/layout invariants, worker reduction, scenario caps/calibration, and cyclic order |
 | `test_llm_memory_executor.cpp` | `LlmMemoryExecutorTest` | CPU phase/layout dispatch, atomic mapping/table-preparation failure seams, contiguous and paged decode/prefill descriptor materialization, independent checksums, padding canaries, synchronized timing, QoS, cancellation, and fake-kernel validation |
 | `test_llm_memory_runner.cpp` | `LlmMemoryRunnerTest` | Fake-backend lifecycle and generic decode/prefill task seams, lifecycle unsupported/failure handling, common identity/timing/completion/validation acceptance, exact calibration/single-unit/freeze/frozen-warmup order, cyclic measurements, status/counter/aggregate semantics, interruption, release/checkpoint precedence, auxiliary budgeting, and runner exception boundaries |
 | `test_llm_memory_json.cpp` | `LlmMemoryJsonTest` | Schema-2 identities for all eight profiles, canonical plan references, expected/actual checksums, named validation, accepted populations, raw Mach timing, build manifest, nullable geometry/workers, paged and Metal resource/grid evidence, output-peak estimates, acceptance, environment, and bounded checkpoint observations |
 | `test_llm_metal_checksum.mm` | `LlmMetalChecksumHelperIntegrationTest` | Real-device checks of shared MSL affine checksum collision boundaries and six-component modulo reduction across partial SIMD/threadgroups |
 | `test_llm_memory_output.cpp` | `LlmMemoryOutputTest` | Exact decode/prefill, contiguous/paged, and Metal device/resource/table/task K/V-write/lookup/padding plus cyclic owner/per-threadgroup cost formatting, interpretation text, and deduplicated warnings |
 | `test_llm_memory_kernels.cpp` | `LlmMemoryKernelIntegrationTest` | Real contiguous/paged decode/prefill ARM64 descriptor/kernel scenarios, paged tails/lookups, full-prompt writes, exact partial tiled scans, checksum, padding, worker-count, and AAPCS64 coverage |
-| `test_llm_metal_backend.cpp` | `LlmMetalBackendTest`, `LlmMetalBackendIntegrationTest`, `LlmMetalBackendFailureInjectionIntegrationTest` | Pure capability, contiguous/paged segmentation, argument-buffer, decode/prefill ABIs, source-hash and prefill loop-order audit, phase-aware checksum/reason mapping, paged-prefill `N+2*M`, cyclic no-duplication ownership and per-threadgroup accounted-byte evidence, plus real-device selected-profile compilation, Tier-2 slots, private table/K/V initialization, exact-tail decode, full-prompt/tiled-prefix prefill, final-ordinal write/padding detection, interruption/failure cleanup, and idempotent release |
+| `test_llm_metal_backend.cpp` | `LlmMetalBackendTest`, `LlmMetalBackendIntegrationTest`, `LlmMetalBackendFailureInjectionIntegrationTest` | Pure capability, contiguous/paged segmentation, argument-buffer, decode/prefill ABIs, source provenance and prefill loop-order audit, phase-aware checksum/reason mapping, paged-prefill `N+2*M`, cyclic no-duplication ownership and per-threadgroup accounted-byte evidence, plus real-device selected-profile compilation, Tier-2 slots, private table/K/V initialization, exact-tail decode, full-prompt/tiled-prefix prefill, final-ordinal write/padding detection, interruption/failure cleanup, and idempotent release |
 | `test_buffer_manager.cpp` | `BufferManagerTest` | Pattern mapping policy, atomic allocation cleanup, initialized content, validation, and peak accounting |
 | `test_benchmark_executor.cpp` | `BenchmarkExecutorTest` | Injected phase/chain failures, continuous latency sampling, and hardware executor contracts |
 | `test_benchmark_runner.cpp` | `BenchmarkStatisticsCollectorTest`, `BenchmarkRunnerTest` | Status-bearing aggregation, schema-3 retained snapshots, checkpointing, interruption, and runner exception/failure seams |
@@ -403,7 +402,7 @@ installed. All test `.cpp` and `.mm` files are picked up automatically by the Ma
 | `test_llm_result_verifier.py` | Independent LLM verifier | Eight current profiles, arithmetic goldens, status/population and semantic mutations, bounded input and build provenance |
 | `test_script_examples.py` | Python script-example entry test | Compatible standard schema-3/methodology inputs, provenance-version handling, completion and field-shape checks, and metric extraction for the bundled standard-memory examples, including the optional jq-backed shell path when jq is installed |
 | `test_json_output_session.cpp` | `JsonOutputTargetTest`, `JsonOutputSessionTest` | Exact sentinel/path classification, lazy checkpoint dispatch, atomic-file parity, stdout routing/restoration, and stream failure containment |
-| `test_json_utils.cpp` | `JsonUtilsTest`, `JsonFileWriterTest` | JSON parse/statistics and atomic writer success/failure contracts |
+| `test_json_utils.cpp` | `JsonUtilsTest`, `JsonFileWriterTest` | UTC timestamp/statistics and atomic writer success/failure contracts |
 | `test_output_printer.cpp` | `OutputPrinterTest`, `OutputPrinterCustomCacheUnitsTest` | Status-aware partial output, mode/cache composition, and custom-cache size-unit boundaries |
 | `test_sweep_runner.cpp` | `SweepRunnerTest` | Current-only nested standard schema-3 completion classification, complete/partial/interrupted/failed attempt accounting, schema-1 envelopes, and file/lazy-stdout checkpoint behavior |
 | `test_sweep_utils.cpp` | `SweepUtilsTest` | Shared sweep parsing, empty-dimension behavior, and overflow-safe Cartesian counts |

documents/TECHNICAL_SPECIFICATION.md

@@ -1482,7 +1482,7 @@ Recommended validation commands:
 - GPU help smoke check: `./memory_benchmark --gpu-bandwidth --help`
 - LLM help smoke check: `./memory_benchmark --llm-memory --help`
 - Deterministic LLM tests:
-  `./test_runner '--gtest_filter=ModeSelectorTest.*:LlmMemoryContractTest.*:LlmMemoryConfigTest.*:LlmMemoryWorkPlanTest.*:LlmMemoryExecutorTest.*:LlmMemoryRunnerTest.*:LlmMemoryJsonTest.*:LlmMemoryOutputTest.*:MessagesTest.*'`
+  `./test_runner '--gtest_filter=ModeSelectorTest.*:LlmMemoryConfigTest.*:LlmMemoryWorkPlanTest.*:LlmMemoryExecutorTest.*:LlmMemoryRunnerTest.*:LlmMemoryJsonTest.*:LlmMemoryOutputTest.*:Messages*Test.*'`
 - Real LLM ARM64 and bounded executable transport contracts:
   `./test_runner '--gtest_filter=*LlmMemory*Integration*:*ExecutableCliIntegration*'`
 - LLM Metal decode and prefill for contiguous and paged KV:

src/benchmark/benchmark_executor.cpp

@@ -287,7 +287,6 @@ int prepare_cache_latency_buffers(BenchmarkConfig& config, BenchmarkBuffers& buf
                                config.custom_buffer_size,
                                config.latency_stride_bytes,
                                config.latency_tlb_locality_bytes,
-                               nullptr,
                                config.latency_chain_mode,
                                derive_benchmark_seed(config.benchmark_seed,
                                                      kSeedDomainCustomLatency));
@@ -303,7 +302,6 @@ int prepare_cache_latency_buffers(BenchmarkConfig& config, BenchmarkBuffers& buf
                             config.l1_buffer_size,
                             config.latency_stride_bytes,
                             config.latency_tlb_locality_bytes,
-                            nullptr,
                             config.latency_chain_mode,
                             derive_benchmark_seed(config.benchmark_seed,
                                                   kSeedDomainL1Latency)) != EXIT_SUCCESS) {
@@ -321,7 +319,6 @@ int prepare_cache_latency_buffers(BenchmarkConfig& config, BenchmarkBuffers& buf
                             config.l2_buffer_size,
                             config.latency_stride_bytes,
                             config.latency_tlb_locality_bytes,
-                            nullptr,
                             config.latency_chain_mode,
                             derive_benchmark_seed(config.benchmark_seed,
                                                   kSeedDomainL2Latency)) != EXIT_SUCCESS) {
@@ -352,7 +349,6 @@ int prepare_main_memory_latency_buffer(BenchmarkConfig& config, BenchmarkBuffers
                              config.buffer_size,
                              config.latency_stride_bytes,
                              config.latency_tlb_locality_bytes,
-                             nullptr,
                              config.latency_chain_mode,
                              derive_benchmark_seed(config.benchmark_seed,
                                                    kSeedDomainMainLatency));
@@ -443,8 +439,7 @@ void run_paired_locality_comparison(void* buffer,
     const LatencyChainMode mode = use_locality
                                       ? LatencyChainMode::RandomInBoxRandomBox
                                       : LatencyChainMode::GlobalRandom;
-    if (setup_latency_chain(buffer, buffer_size, stride_bytes, locality_bytes,
-                            nullptr, mode, seed) != EXIT_SUCCESS) {
+    if (setup_latency_chain(buffer, buffer_size, stride_bytes, locality_bytes, mode, seed) != EXIT_SUCCESS) {
       return false;
     }
     show_progress();
@@ -1047,14 +1042,6 @@ BenchmarkResults run_single_benchmark_loop(BenchmarkConfig& config,
         throw std::runtime_error(
             Messages::benchmark_reason_prepare_failed(enabled_phase.name));
       }
-      if (test_hooks != nullptr && test_hooks->fail_latency_chain_setup &&
-          (enabled_phase.phase == Phase::CacheLatency ||
-           enabled_phase.phase == Phase::MainLatency) &&
-          test_hooks->fail_latency_chain_setup(enabled_phase.name)) {
-        throw std::runtime_error(
-            Messages::benchmark_reason_latency_chain_setup_failed(
-                enabled_phase.name));
-      }
       switch (enabled_phase.phase) {
         case Phase::MainBandwidth: {
           BenchmarkBuffers phase_buffers;

src/benchmark/benchmark_executor.h

@@ -38,7 +38,6 @@ struct BenchmarkExecutionState;
 /** @brief Optional kernel-free fault seams for phase preparation tests. */
 struct BenchmarkExecutorTestHooks {
   std::function<bool(const std::string&)> fail_phase_preparation;
-  std::function<bool(const std::string&)> fail_latency_chain_setup;
 };
 
 /**

src/core/config/constants.h

@@ -259,8 +259,6 @@ namespace Constants {
       "llm-kv-layout-geometry-v1";
   constexpr const char* LLM_KV_LAYOUT_PLAN_IDENTITY_VERSION =
       "llm-kv-layout-plan-v1";
-  constexpr const char* LLM_PAGED_DECODE_WORKLOAD_IDENTITY_VERSION =
-      "llm-paged-decode-workload-v1";
   constexpr const char* LLM_PAGED_CPU_EXECUTION_IDENTITY_VERSION =
       "llm-paged-cpu-execution-v1";
   constexpr const char* LLM_PAGED_METAL_EXECUTION_IDENTITY_VERSION =