Remove duplicate fresh-index embedding pass

benchmarks/README.md

@@ -194,6 +194,50 @@ Writes to `benchmarks/results/speed-<sha12>.json`.
 
 </details>
 
+<details>
+<summary>Cold index phase profiler</summary>
+
+Profile fresh `create_index_from_path` calls without enabling production telemetry:
+
+```bash
+uv run python -m benchmarks.profile_cold_index \
+  --corpus-path /path/to/source-tree \
+  --model-path /path/to/model \
+  --repetitions 5 \
+  --label experiment-name \
+  --revision revision-id \
+  --output /path/to/profile.json
+```
+
+The model is loaded before measurement; every repetition passes `previous=None`. Cold means fresh index state, not a
+fresh process or cold parser/filesystem caches. `records` contains raw per-run nanosecond timings, counts, call-size
+arrays, index shape/vector bytes, and process peak RSS. `summary` recursively reports median/min/max for numeric fields
+and `embedding_invariants` states whether each duplicate-pass check held across every repetition.
+
+`total.instrumented_wall_ns` and `total.instrumented_process_cpu_ns` surround `create_index_from_path`, including all
+wrapper and recorder overhead. They must not be compared with an uninstrumented benchmark; measure that total
+separately. Phase `wall_ns` is inclusive, `exclusive_wall_ns` removes nested measured boundaries, and
+`overlap.nested_wall_ns` names each parent/child overlap. Inclusive parent and child times must not be summed.
+
+For each phase, `calls` counts boundary entries and `items` counts outputs appropriate to that boundary.
+`file_walk_iterator` includes traversal work and its internal filesystem checks. `file_status_checks` includes file
+eligibility stats and sub-128-byte emptiness probes; its `items` count is checked files, not individual stat calls.
+`source_reads` covers only the subsequent production source read;
+`counts.source_read_bytes` is the UTF-8 size of its returned text and is counted after the timer. Direct manifest mtime
+stats are unattributed. `call_sizes` plus count/total/median/min/max are emitted for BM25 updates, `embed_chunks`,
+`StaticModel.encode`, and Model2Vec tokenizer batches. `index` reports files, chunks, dimensions, and vector bytes.
+
+`static_model_encode` is the exact `StaticModel.encode` API boundary, not pure native model time.
+`model2vec_tokenization` is nested within it; `model2vec_lookup_mean_stack_normalization` is the derived remainder
+(`encode - tokenize`) and includes Python overhead. `unattributed_including_profiler_overhead` is the root exclusive
+remainder: unwrapped index orchestration, vector/BM25 finalization, manifest mtime stats, and profiler bookkeeping.
+
+`memory.process_peak_rss_bytes` is a monotonic process-lifetime high-water mark. It includes the preloaded model and
+all earlier repetitions, so it is not a per-repetition allocation delta. The recorder is single-thread-only; a future
+parallel indexer requires thread-local or synchronized instrumentation.
+
+</details>
+
 <details>
 <summary>Ablations</summary>