Function is_internal_counter 

pub fn is_internal_counter(name: &str) -> bool

SRD-100 P2 — the per-phase status builder, re-exported for the display consumer (nmbrs-tui) that now folds active_phases and renders each phase’s status line itself (the module otherwise stays crate-private). SRD-100 P2 — the per-phase status builder, re-exported for the display consumer (nmbrs-tui) that now folds active_phases and renders each phase’s status line itself (the module otherwise stays crate-private). SRD-100 P2 — the per-phase status builder, re-exported for the display consumer (nmbrs-tui) that now folds active_phases and renders each phase’s status line itself (the module otherwise stays crate-private). Average batch size for the rows/batch: chip — rows written per successful batch op.

Prefers rows_inserted / batch_writes, where batch_writes is the number of ops that actually wrote ≥1 row (published by the CQL batch dispensers alongside rows_inserted). This is the true per-op stride: retried, failed, and non-inserting ops never touch batch_writes, so the denominator can’t drift the way stanzas_total (one inc per op attempt, regardless of success/failure/type) does.

Falls back to rows_inserted / stanzas_total when no batch_writes counter is present — non-CQL or older adapter paths that never learned to publish it. Returns None when no batched write was observed (average would be ≤1 row/op, i.e. not a batch).

Kept as a free function (not inline at the call site) so the three display paths that render this chip — the inline refresh here and the two TUI progress-thread snapshots in executor.rs — share one formula, and so the formula is unit-testable in isolation. Whether a dispenser status counter is INTERNAL — published only to feed a derived display metric, never rendered as its own <name>/s throughput chip. The convention is a leading underscore: _batch_writes backs the rows/batch average (see [rows_per_batch]) but must not clutter the operator-facing chip row as _batch_writes/s. Every surface that turns a dispenser counter into a chip filters on this one predicate, so the convention has a single point of truth. find_counter-style lookups pass the underscore name explicitly, so the counter stays available to the derived metric it exists for.