chore(vendor): split remaining generated and data churn out of PR #707

benchmark_data/README.md

@@ -0,0 +1,10 @@
+# Benchmark data and provenance
+
+This directory contains benchmark workloads, fixtures, harness scripts, and
+captured result evidence. Cargo benchmark source remains in `benches/` and in
+crate-local `benches/` directories.
+
+Current scripts and manifests use `benchmark_data/` paths. Captured historical
+result documents may retain a `benchmarks/` path when that value records the
+repository layout at capture time; those values are evidence, not live path
+references.

benchmark_data/claude-observation/README.md

@@ -0,0 +1,128 @@
+# Provider-observation pipeline benchmark
+
+> **Historical evidence only.** Preserve the authentic provider fixtures,
+> measurements, and provenance in this directory. Current requirements come
+> only from the `docs/plans/tracedecay-v2/` hierarchy; exact commands, counts,
+> snapshots, receipts, attestations, PR packets, and gate fields below are not
+> rebuild instructions. Validate current provider behavior directly.
+
+Archival result artifacts keep the workload id they were sealed under for
+provenance. Workload schema 3 is the historical multi-provider Linux
+benchmark; the checked-in results are descriptive measurements, not PR
+acceptance authority.
+
+The versioned [workload](workload-v1.json) runs production scan/parse,
+normalization, sanitizer, authoritative commit, projection/V1 fold, bounded
+replay, and exact repeat-ingest no-op paths for claude, codex, cursor, hermes,
+kiro, cline, roo-code, and kilo. Each provider baseline machine-asserts parse,
+normalize, sanitize, commit, replay, duplicate-noop, projection, backlog,
+fairness, and peak-resource checks. The measured loop executes one provider
+turn per round in catalog order, records every turn, and validates an
+eight-provider maximum turn distance. Each provider input is a bounded copy of
+the checked-in native fixtures named in the workload; unique identities and
+secret-shaped canaries are injected into those native shapes and the canaries
+must be absent from durable observations. The historical result embeds a versioned
+`provider-observation-performance-result-v1` containing per-provider raw
+parse/commit/replay/pipeline/no-op samples, p50/p95/p99 distributions,
+throughput, CPU, process write I/O, database growth, peak RSS, bounded replay backlog, and
+observation-count deltas. The pipeline is the real production
+parse/normalize/sanitize/commit/project/replay path. The separate parse
+distribution measures native provider-format decoding over the exact bounded
+input, commit measures the full production adapter through authoritative commit
+and projection, and replay measures the bounded authoritative-store query.
+Scopes are embedded beside each distribution so the commit sample is not
+misrepresented as excluding the adapter's parse and normalization work.
+
+The provider catalog remains schema v1. Its measurement block is an additive
+optional field for wire compatibility, while executable validation requires
+that block for every supported provider and rejects pending paths. This avoids
+an unnecessary incompatible schema-v2 claim. A normal test deserializes the
+complete manifest with unknown fields denied, executes every production adapter
+once, and proves each repeat is a durable no-op.
+
+The current measurement includes the nested
+`provider_observation_performance` result and `hook_telemetry_readiness`
+diagnostic. [evidence-index.json](evidence-index.json) identifies
+[result-2026-07-26-dc17dd73.json](result-2026-07-26-dc17dd73.json) as the
+legacy `current_acceptance`; that field is deprecated and grants no authority.
+Earlier artifacts remain `historical_stale`.
+
+The measurement embeds hook telemetry readiness as
+`hook-telemetry-baseline-readiness-v1`, not as a measured baseline. It reads the
+redacted fixtures under `tests/fixtures/host_events` directly, records each
+fixture path and SHA-256 identity, and computes compact canonical request-byte
+samples. It consumes `crate::hooks::host_hook_telemetry_contract` and
+`crate::hooks::measure_host_event_payload_bytes`; it defines no second runtime
+telemetry schema. Current production telemetry exposes hook wall time, daemon
+RTT, host-event/IPC byte counts, timeout state, and disposition per invocation
+on instrumented hosts; Hermes coverage remains partial. Checked-in distributions
+and aggregate timeout/disposition counts remain explicitly unavailable rather
+than being represented as zero.
+
+The legacy schema records real SHA-256 identities for workload members and
+native fixtures plus the host target, Rust/Cargo versions, kernel, and hardware.
+Those source/content digests remain useful provenance. Do not recreate the
+former clean-checkout archive, tracked-source snapshot, compiler-input
+attestation, or evidence-only commit workflow. Run ordinary Cargo against the
+working checkout and report the source revision and dirty state truthfully.
+
+Linux `/proc` is the explicit measurement platform contract. Preflight requires
+all measured interfaces, a successful write of `5` to
+`/proc/self/clear_refs`, and a nonzero `getconf CLK_TCK`, before warmup begins.
+CPU, process-write, storage-growth, and peak-RSS provider fields therefore
+contain measurements only for successful Linux evidence runs. Unsupported
+platforms reject at preflight; the harness never substitutes numeric zero for
+an unavailable counter. Zero is reserved for an available counter that
+actually measured no work, including the strict repeat-ingest no-op assertion.
+The module still compiles on non-Linux targets; an attempted measurement there
+rejects the unsupported platform at preflight. CPU identity accepts common x86,
+ARM, POWER, and other Linux
+`/proc/cpuinfo` labels.
+
+Every replayed authoritative payload is checked for canary removal and a
+payload-bound sanitization receipt, and every folded V1 message is checked for
+exact identity, role, text, and canary absence. These assertions and V1 point reads run after
+each measured phase, so correctness verification is not charged to latency,
+CPU, I/O, or storage-growth measurements. The run also requires zero legacy
+transcript writes: the observation projector is the only V1 message writer,
+and compatibility transcript counters report those projector outputs.
+The timed no-op retry replays after the durable end cursor and must return zero
+new observations; a full replay verifies unchanged cardinality afterward.
+
+The current [measurement result](result-2026-07-26-dc17dd73.json) records
+source revision `dc17dd731e97f9262e570afd9e7ec1602d8af99e` with 3 warmups and
+30 independent measured repetitions of 64 records (30 × 64 = 1,920 records).
+The raw artifact records the Linux kernel, CPU, memory, Rust/Cargo toolchains,
+every repetition, and the nearest-rank/sample-standard-deviation method.
+
+- Pipeline batch latency: p50 2,341,276,568 ns; p95 2,527,159,267 ns; p99
+  2,708,492,865 ns.
+- Pipeline throughput: 27.11224430907219 records/s.
+- Timed pipeline CPU: 69,110 ms; peak RSS: 153,276 KiB.
+- Timed process write I/O: 1,551,220,736 bytes; SQLite database/WAL/SHM growth:
+  1,122,556,640 bytes across the 30 independent databases.
+- Exact no-op retry plus bounded replay: p50 1,185,536 ns; p95 1,504,741 ns;
+  p99 1,558,623 ns; 40 ms CPU total; zero process write bytes, database growth,
+  observation-count delta, and coordinator work counters.
+- Round-robin fairness: maximum provider turn distance 8.
+
+The earlier [measurement result](result-2026-07-16-00d3d73a.json) remains
+`historical_stale` after the checked-in workload identity changed.
+
+The earlier [measurement result](result-2026-07-16-8d53b4a9.json) remains
+`historical_stale` after the PR6 provider-ingestion correctness changes.
+
+The earlier [measurement result](result-2026-07-15-0c289212.json) remains
+`historical_stale` because it predates workload schema 3.
+
+The [historical result](result-2026-07-15-b05b4cd5.json) was captured from clean
+commit `b05b4cd570ab8e3385604c0fef31902fdc3f1e8b`.
+
+> **Historical/stale evidence:** this result predates schema 2 provenance and
+> complete workload validation. Its JSON carries
+> `"evidence_status": "historical_stale"` and is not current product evidence.
+> Retain it only for provenance.
+
+```console
+scripts/run-claude-observation-benchmark.sh
+```

benchmark_data/claude-observation/evidence-index.json

@@ -0,0 +1,11 @@
+{
+  "schema_version": 1,
+  "current_acceptance": null,
+  "historical_stale": [
+    "result-2026-07-15-0c289212.json",
+    "result-2026-07-15-b05b4cd5.json",
+    "result-2026-07-16-00d3d73a.json",
+    "result-2026-07-16-8d53b4a9.json",
+    "result-2026-07-26-dc17dd73.json"
+  ]
+}