v0.23.0

Minor Changes

• 32df428: Breaking (query blame --format json): each read-site row is now a structured location
  ({ source, line?, column?, ... }) matching the human table's columns, instead of a single
  file:line:col string under at. Split at yourself only if you still need the joined form.

  Add bot-wall detection: when wpd's own navigation (the built-in --url load flow, a --url host page)
  lands on a bot-challenge interstitial (Cloudflare, DataDome, hCaptcha, PerimeterX, Arkose), record
  refuses before measuring — non-zero exit, an evidence-listed error, and a <recording>.wall.png
  screenshot — rather than reporting the challenge page as the site. Detection is conservative (rendered
  interstitial only, never a captcha script a form embeds). --allow-bot-wall measures it anyway, with a
  loud note.

  Add siteRelation (same-origin | same-site | cross-site) on query cpu origin buckets of a
  --url run, via the public-suffix list. It is a URL-mechanical fact, never an ownership or
  "third-party" claim (a cross-site CDN can be first-party-owned).

  The built-in --url flow now names the failure class (navigation timeout, HTTP/2 reset,
  context-destroyed) and points at the driver-module escape hatch: wpd retries its own machinery's races,
  never the site's refusals.

• ac00852: Record Chrome's own soft-navigation verdict on a driver step, beside the url+timeOrigin classifier. On
  a Chrome that ships the Soft Navigations API (151+, default-on) a step carries engineSoftNav
  (count, navigationTypes, and the numeric ids) from an in-page soft-navigation observer. It is
  opportunistic: wpd never forces --enable-features, so an older Chrome or Firefox records nothing, and
  absence is never a fabricated 0.

  query span reconciles the two verdicts: where the classifier reads a step "soft" but the engine fired
  no entry (a programmatic history change, an untrusted click, or no qualifying paint), it notes both and
  picks no winner, so a route the engine's metrics miss is visible rather than hidden.

• bcb0a9b: Add --alloc: a node-lane allocation-attribution capture mode. wpd record <module> --target node --alloc runs V8's heap sampler (GC-inclusive) around your run() loop and attributes allocated bytes
  to source/package, answering "which dependency allocates". Read it with the new query alloc --by package|file|function. It is a dedicated mode with the CPU sampler OFF (a co-riding heap sampler
  inflates CPU self-time), so an --alloc recording carries no CPU model; query cpu/cpu-diff on one
  point you at query alloc. Byte shares/ratios are trustworthy (~5%); the absolute total is directional
  (~10-20%).

• 875c448: Boot LCP is now per-iteration sampled: under --iterations N the load step's lcp grows
  perIteration (the render-time series, null for an iteration that fired no entry, never 0) and
  stats (min/median/max), the same shape wall carries, so a run-to-run LCP swing is visible instead
  of hidden behind one number. query span prints the spread; the identity fields stay a real sample.

  Add per-step CLS: a driver step carries layoutShift (Chrome only) — the spec session-window maximum
  (session windows gap-capped at 1s / window-capped at 5s, hadRecentInput shifts excluded), not a raw
  sum, with the top shifting elements attributed (tag#id, rect deltas). Scoped to the step's own
  window; Firefox has no layout-shift entry type, so it is absent there, never a fake 0. Both are
  additive fields (schema stays 4).

Patch Changes

• e745844: Move the bundled browser to Chrome 151 (Puppeteer 25.4.0), so CI, e2e, and a fresh install all run
  one browser. Re-probed the load-bearing headless facts on 151: the one-frame floor (16.7 ms / 60 Hz)
  is unchanged; the GPU frame-sink stall no longer reproduces (its --in-process-gpu forcing lever now
  produces frames cleanly), so the --disable-gpu default stays as belt-and-braces; boot-LCP delivery
  still recovers within its bounded budget, with no missing-entry or 60 s-startTime anomaly; and a
  cross-origin LCP without Timing-Allow-Origin now reports a coarsened renderTime (more data, not
  wrong data). Soft-navigation entry types (soft-navigation, interaction-contentful-paint) are
  present by default in the measured browser.
• db12e7a: Lower the cross-process split-detection floor so a lighter second navigation is not silently
  uncounted. A run that navigates across renderer processes now sets meta.mainThread.split (so
  assert / diff --fail-on-regression refuse count gates) once the second navigation renders at least
  5% of the busiest thread's layout/paint -- the same husk share the re-anchor uses. A second navigation
  doing 5-24% of the first page's work previously left split false, so assert --max-layouts gated
  green on the first page's counts alone. Keep each run to one navigation for counts that cover all of it.