v0.3.3

Tracks alpenglow ag-v0.3.2. Post-release polish on standstill
reporting, runner text-mode verdicts, and Time Series presentation,
plus a new transaction-pressure card.

Added

• Transaction-pressure card on the Time Series tab. Series is
  Σ signature_count / bucket_seconds from bank frozen log
  lines, rendered as tx/s (bucket-size-independent so 10m and
  1h zoom levels read on the same scale). Stats line surfaces
  min / avg / peak per second alongside an overall
  tx/block (avg) derived from Σ signature_sum / Σ blocks.
• Metric::with_subtitle on the Time Series renderer for inline
  secondary stats on existing cards (no new grid cell required).
• Kind::Rate variant for per-second metrics. Distinct from
  Kind::Count because summing a rate across buckets is
  meaningless; the stats line therefore drops the total and
  /bkt fields for Rate cards.

Changed

• StandstillObserved alerts dedup by anchor slot: repeated
  StandstillExtending events at the same at_slot merge into
  one alert with count (firings) and last_at (most recent
  extension). Index lives on OverallStats.standstill_alert_indices
  (transient). Previously a 2.5h standstill produced one alert
  per extension batch (hundreds of rows).
• Standstill alert detail pane shows anchor / firings / first at / last at / span instead of a single timestamp.
• Sparkline x-axis caption marked (autoscale) so readers know the
  range is data-driven, not a fixed window.
• Headline-health canonical-skip line shortened from
  (N canonical of M vote-skips · K indeterm) to
  (N of M · K indeterm) to fit single-line width.
• Local-leader summary row drops the
  (4-slot bursts, per NUM_CONSECUTIVE_LEADER_SLOTS) qualifier.
• runner.rs text-mode health lines take separate ok_note and
  bad_note strings. Previously both paths reused one note,
  producing contradictory output like [✗] no liveness issues.
  Standstill bad-path verdicts now read STANDSTILL OBSERVED /
  resume activity instead of no liveness issues.
• Runner crashed-leader count is standstill-aware when
  standstill_ranges exist: subtracts TCLs that fired inside a
  standstill window from the headline figure.
• cargo xtask lint-prod mirrors the CI strict clippy gate
  exactly: cargo clippy --workspace --all-targets --locked -- -D warnings. Previously it omitted --all-targets,
  --locked, and -D warnings, allowing local "pass" while
  CI failed.

Fixed

• Tx-pressure stats line previously claimed total N peak K/bkt avg J/bkt for a per-second series. total was the sum of
  rates (meaningless); the /bkt suffix lied about the unit
  (values were per-second). Fixed by routing the card through
  Kind::Rate.
• clippy::derivable_impls on SkipClassification: replaced
  manual Default impl with #[derive(Default)] +
  #[default] on NotSkipped.
• clippy::manual_checked_ops (Rust 1.96+) on the
  avg_tx_per_block divide-by-zero guard: replaced
  if d > 0 { a / d } else { 0 } with
  a.checked_div(d).unwrap_or(0).

Toolchain

• Added rust-toolchain.toml pinning the project to Rust
  1.96.0. .github/workflows/{ci,release}.yml install the
  same version via dtolnay/rust-toolchain@master with
  explicit toolchain: 1.96.0. Previously local and CI both
  said "stable" but tracked different versions, so new lint
  categories (e.g. manual_checked_ops in 1.96) appeared at
  push time. Bumping Rust now means editing one file plus the
  two workflow inputs in lockstep.