feat(tui): split each token-meter band into new and cache-read work

[edwin-zvs]

What

A turn's prompt side is mostly context the provider already holds, so a meter column that reads as 90k of work can be 80k of re-sent conversation. The meter had no way to say which — cached input was deliberately excluded from the total to avoid double-counting, and then discarded.

Now it rides along. TokenSample gains a cached field: a subset of tokens (so tokens - cached is the new work), #[serde(default)] so records written before it existed load as 0. It's populated by both the live usage path and the boot-time transcript scan, so a restarted client draws the same columns it had before.

Each model's band then splits in two — its new work at the base, the cache-served remainder directly above it, in the same hue pushed toward black (blend_color(Black, base, 0.45)).

• Bar height is unchanged: it stays billed volume. Cached is shading, not an addend.
• Legend rates are unchanged: still tokens-per-second of billed volume.
• Hover names the exact cached figure per model (opus 90.0k (80.0k cached)), which is the only place the shaded part gets a number.
• A zero part draws no band at all, and a harness over-reporting cached is clamped to the sample so the subtraction can't underflow.

The recessed shade has to be a color, not Modifier::DIM: one terminal cell can hold a model's own new/cached boundary, and there the two parts are the glyph's foreground and its background — an attribute would dim both or neither. ColumnCell/column_cells/stacked_eighths are now generic over the band key so the same machinery draws a same-model boundary exactly like a cross-model one.

Also: more margin between legend columns (1 → 3). A rate sits against its cell's right edge and the next cell opens with a dot, so a single space of separation read as one run of text rather than as columns. The gap is a named constant folded into the cell width, so the rate anchor, the column count, and the inter-column spacing stay consistent.

Spec

specs/0167-fleet-token-meter.md gains "Cached input is a subset, and it shades a band rather than growing it" — the subset contract, why height stays billed volume, why same hue and why a color rather than an attribute, and what a harness reporting no cache figure looks like. The legend-grid section picks up the per-cell margin, and "The stack must survive the cell grid" is reworded from series boundary to band boundary to cover the same-model case.

Tests

New unit coverage for: the split not changing column height, each model keeping its two parts adjacent, a zero part producing no band, over-reported cached being clamped, seeded history keeping its cached share, the cached color being darker on every channel while holding hue, and a same-model boundary encoding as fg-over-bg. Legend tests assert against the gap constant rather than a literal.

Binaries

Touches crates/cli, crates/daemon, crates/protocol → the user-visible change is in construct.