Bench pr 35215 #35739

Adds build configuration, benchmarking CI workflows, Python
dependencies, plotting tools, and documentation for benchcoin.

Co-authored-by: David Gumberg <davidzgumberg@gmail.com>
Co-authored-by: Lőrinc <pap.lorinc@gmail.com>

- Fix empty chart: use get_chart_data() instead of to_dict() so JS
  filters can match config strings ("450", "32000") instead of objects
- Capture machine specs on self-hosted runner during build job and pass
  via --machine-specs flag to nightly append, instead of detecting on
  the ubuntu-latest publish runner

Run LogParser + PlotGenerator from bench/analyze.py during artifact
copying to produce static PNG charts from debug.log files. This
pre-generates the same 11 chart types that were previously rendered
client-side via JavaScript.

Changes to report.py:
- Import HAS_MATPLOTLIB, LogParser, PlotGenerator from bench.analyze
- _copy_network_artifacts: generate plots after each debug.log with
  "{network}-{name}" prefix (e.g. "450-uninstrumented-pr")
- _copy_artifacts: generate plots for single-directory mode, including
  when input_dir == output_dir
- _prepare_graphs_data: add "plots" key with relative paths to PNGs
- generate(): reorder to copy artifacts before HTML rendering so
  _prepare_graphs_data can find the generated plot files

Plot generation is guarded by HAS_MATPLOTLIB for graceful fallback
when matplotlib is unavailable.

The pr-report.html template previously included debug-log-charts.html
which fetched multi-hundred-MB debug.log.gz files in the browser,
decompressed them with pako.js, parsed every line, and rendered 11
Plotly charts client-side. This made report pages unresponsive.

Now that report.py pre-generates the charts as static PNGs:
- pr-report.html: replace the debug-log-charts.html include with an
  img loop over graph.plots, using loading="lazy"
- debug-log-charts.html: delete (344 lines of client-side JS)
- base.html: remove pako.js and Plotly CDN scripts (both are
  independently included by pr-chart.html and nightly-chart.html
  via their own script tags)

The debug.log download link is preserved.

Rewrite to document the TOML config + matrix entry workflow,
removing stale references to the old two-commit comparison CLI,
--datadir requirement, profiles, and BENCH_DATADIR env var.

Debug logs were consuming 388MB on gh-pages. They are already uploaded
as CI artifacts with 90-day retention during benchmark runs.

- Remove gzip compression and copying of debug logs in report generation
- Remove debug log extraction in publish-results workflow
- Replace per-graph "Download debug.log" links with a single link to
  the CI run page where artifacts can be downloaded
- Keep matplotlib plot generation from debug logs (plots are still
  generated during report phase, just the raw logs aren't published)

The PR comment with result links was posted before GitHub Pages
finished deploying, leading to broken links. Add a wait-for-pages
job that polls for the pages-build-deployment run matching our
exact gh-pages commit, then blocks until it completes.

Previously, prune=10000 was causing flushes of the UTXO set when block
pruning was taking please, resulting in logs like:

❯ zcat 32000-instrumented-pr-debug.log.gz | rg UTXO
2026-02-12T07:22:57Z * Using 31990.0 MiB for in-memory UTXO set (plus up to 286.1 MiB of unused mempool space)
2026-02-12T07:28:51Z [warning] Flushing large (2 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:33:10Z [warning] Flushing large (3 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:37:23Z [warning] Flushing large (4 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:42:03Z [warning] Flushing large (4 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:46:34Z [warning] Flushing large (5 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:51:10Z [warning] Flushing large (6 GiB) UTXO set to disk, it may take several minutes
2026-02-12T07:55:57Z [warning] Flushing large (7 GiB) UTXO set to disk, it may take several minutes
2026-02-12T08:00:35Z [warning] Flushing large (8 GiB) UTXO set to disk, it may take several minutes
2026-02-12T08:05:16Z [warning] Flushing large (8 GiB) UTXO set to disk, it may take several minutes
2026-02-12T08:10:00Z [warning] Flushing large (8 GiB) UTXO set to disk, it may take several minutes
2026-02-12T08:14:36Z [warning] Flushing large (8 GiB) UTXO set to disk, it may take several minutes
2026-02-12T08:16:47Z [warning] Flushing large (8 GiB) UTXO set to disk, it may take several minutes

and generally interrupting benchmarking. Remove this effect by setting
prune to such a high value it will never trigger.

Prune is **required** to permit us to continue syncing from a pruned
datadir.

ec9395419a removed debug log copying to stop publishing raw logs to
gh-pages (388MB). But it also removed the extraction step that makes
debug logs available during report generation, so matplotlib had no
input files and PNG charts silently stopped appearing in PR reports.

Restore the copy of debug-logs-${network} artifacts into the results
directory before report generation. The raw logs are still not committed
to gh-pages — only the small pre-rendered PNGs in plots/ are.

Replace the matrix strategy with explicit sequential jobs so the
450 benchmark always runs first in a consistent cache state, and
the 32000 benchmark always runs second.

Weekly fstrim.timer on the runner caused a ~25% speedup every Monday
(Sunday night run). Running fstrim in the prepare script before each
benchmark ensures consistent write performance regardless of when
the system timer last ran.

Follows the same suid wrapper pattern as drop-caches.

This reverts commit fd8bdf64ec654c3dbc8677a2bfffa005cbac5ed0.

FITRIM ioctl requires the filesystem mount point. Resolve it from
the tmp_datadir path by walking up to the mount boundary.

Manual (workflow_dispatch) runs are now stored separately from scheduled
nightly runs. Scheduled runs still dedup by (date, commit, dbcache) to
handle retries. Manual runs always append, appearing as diamond markers
on the chart alongside the nightly trend line.

Also ruff format.

Manual (workflow_dispatch) runs no longer get a separate "(manual)"
legend entry with diamond markers. They appear as regular points in
the same series trace as scheduled runs.

Adds a separate benchmark job (benchmark-noav) that runs IBD with
-assumevalid=0 to measure full script verification performance.
Uses a dedicated TOML config with uninstrumented-only matrix, and
prefixes artifacts with noav- so the publish workflow can handle
them alongside existing runs.

Move benchmark shape into declarative experiment manifests so subjects, profiles, and derived comparisons can change without invasive workflow edits. This gives PR, nightly, and differential flamegraph runs the same high-level interface while keeping the existing build and benchmark phases as implementation details.

Record perf artifacts from instrumented runs and derive flamegraph outputs from the hyperfine cleanup hook. This keeps artifact conversion out of the measured command and gives comparisons a stable raw profile input.

Cut PR and nightly workflows over to experiment output trees and drop active no-assume-valid benchmarking. Historical noav result parsing is left in report generation so old published pages continue to render.

Make experiment manifests the only public benchmark shape. The old run command, BenchmarkConfig module, and bench/configs TOML files duplicated profile semantics now owned by Experiment.

Introduce RunSpec as the internal contract consumed by BenchmarkPhase. Nightly history now records config snapshots from experiment manifests instead of loading legacy benchmark configs.

Centralize experiment artifact layout and write artifacts.json so workflows and derivations no longer need to duplicate run directory conventions.

Update PR publish and nightly append workflows to read result, debug, flamegraph, and profile metadata from the artifact manifest instead of hardcoded experiment-output paths.

Make experiment selection produce explicit build, run, and comparison tasks before any execution starts. This keeps validation and filtering separate from the runner and gives dry-run output the same source of truth as real runs.

The runner now consumes the plan directly, which makes profile-scoped comparison derivations explicit and reduces the amount of scheduler logic embedded in execution.

Separate process-level execution settings from benchmark run specs so build, benchmark, and experiment phases no longer depend on the broad Config object.

The CLI still resolves bench.toml, environment variables, and arguments in one place, then hands phases small environment objects with only the flags and paths they need. Build dry-runs now return before saving or restoring git state.

Make CI report generation consume artifacts.json directly instead of rebuilding profile inputs as network-named directories. The report model now carries profile metadata through rendering and combined results output.

This removes the old noassumevalid display path and drops the unused multi-network report CLI surface, leaving single-run reports and experiment-manifest reports as the two supported modes.

Move commit and binary subject handling behind a resolver so the experiment runner executes a plan without also owning subject acquisition policy.

This makes future subject sources, such as prebuilt artifact lookups or named branch refs, a localized change instead of another branch in the runner loop.

Move CI-facing result selection into benchcoin so workflow code no longer has to understand artifact manifest internals, profile naming, or report JSON shape. Recording run config snapshots in artifacts keeps dbcache and instrumentation identity local to the experiment output, which makes publishing and nightly history less brittle.

Report generation and nightly history should not know the raw artifacts.json run schema. A typed catalog interface keeps path resolution and manifest field knowledge local to ArtifactStore, so future artifact layout changes have one place to update.

The nightly benchmark derives the Bitcoin commit by running merge-base against upstream/master. The current benchcoin stack is deeper than the old shallow checkout, so merge-base failed before the experiment could start. Use a bounded but larger fetch depth to keep the job lightweight while covering the current stack.

Use experiment config snapshots when formatting PR report labels and summaries so profile-specific bitcoind arguments are visible in published results. This keeps default nightly-comparable runs on the existing comparison path while allowing exploratory matrices without matching nightly baselines to publish their raw run times.

Plotly can render hover and legend text using trace colors on white surfaces. The previous palette included very light entries, especially the lime series, which made labels hard to read.

Keep the existing ten-slot palette and deterministic color indexes, but replace each slot with a darker same-hue color that clears normal text contrast against white.

The report summary formatter now uses config snapshot labels so PR comments preserve dbcache units and any extra bitcoind arguments consistently.

Refresh the CI interface assertion to match that canonical label instead of the older display-only "450 MB" spelling.

Add deterministic fixed-width and block-sequence SipHash-2-4 vectors before refactoring the implementation.

Validate the shared JSON with `CSipHasher`, the applicable `PresaltedSipHasher` overloads, and the independent Python implementation.
Move randomized generic/fixed comparisons to the integer fuzzer.

The later SipHash-1-3-UJ tests can add expected outputs and comparisons to the same inputs.
Migrating the separate 0-63-byte `siphash_4_2_testvec` incremental-write test to JSON is out of scope.

Co-authored-by: Pieter Wuille <pieter@wuille.net>

Make the new SipHash-1-3-UJ construction smaller and easier to verify by extracting the round, normal-block, and finalization logic duplicated across `CSipHasher` and the fixed-width `PresaltedSipHasher` paths.
The following commits can then focus on the differences from SipHash-2-4.

Co-authored-by: Pieter Wuille <pieter@wuille.net>

Add `SipHasher13UJ`, a block-oriented SipHash-1-3-UJ implementation accepting arbitrarily mixed 64-bit normal blocks and 256-bit jumboblocks.
Use this generic interface as the executable specification for the fixed-width implementation added next.

Co-authored-by: Lőrinc <pap.lorinc@gmail.com>
Co-authored-by: Jean-Philippe Aumasson <jeanphilippe.aumasson@gmail.com>

Add `PresaltedSipHasher13UJ` for one `uint256` jumboblock, optionally followed by one 64-bit normal block.
Implement these paths directly in the header because the new hasher benefits from inlining, unlike the existing SipHash implementation.

Co-authored-by: Lőrinc <pap.lorinc@gmail.com>

Add SipHash-1-3-UJ expected outputs to the shared vectors and check block-encoding invariants and that the standard and UJ finalizers remain distinct.
Compare the generic and fixed-width implementations for every overlapping input and fuzz the same equivalence.
The UJ expected outputs were generated by a reference implementation that Claude Opus 4.8 produced from only the `SipHasher13UJ` class comment.

Co-authored-by: Lőrinc <pap.lorinc@gmail.com>

Add 32- and 36-byte benchmarks for SipHash-2-4 and SipHash-1-3-UJ after establishing their shared correctness coverage.
Keep all four measurements together so both fixed shapes can be compared before changing `SaltedOutpointHasher`.

Co-authored-by: Pieter Wuille <pieter@wuille.net>

Switch `SaltedOutpointHasher` to `PresaltedSipHasher13UJ`, accelerating the coins cache and other in-memory outpoint tables while preserving the public wrapper type.
Its salted outputs are process-local and must not be persisted, serialized, or compared across processes.

Co-authored-by: Pieter Wuille <pieter@wuille.net>