feat(tui): give navigation-work metrics distinct visual sections

[styler-ai]

Why

The token-impact TUI gives File reads avoided a clear structured row, but renders Broad folder walks skipped and Candidate files not opened as dense stacked text and bars below it. In the current normal-width view those two categories are cramped and visually subordinate even though they communicate substantial, distinct impact.

The preferred reference uses a consistent columnar rhythm, visible separation, and a clear label/value/bar hierarchy. Each navigation-work category should receive an equally deliberate treatment so users can scan the dashboard without first decoding which lines belong together.

What Changes

• Present file reads, folder walks, and candidate files as three clearly distinguishable peer sections or rows.
• Use the preferred reference's structured visual language for each category: category identity, exact count, short meaning, labeled proportional bar or bars, and confidence/context where relevant.
• Preserve the two independent folder-walk and candidate denominators:
  • activity share against all reconciled persisted source steps;
  • token-impact share against reconciled tokens avoided.
• Preserve the observed/modelled file-read split and the exact WHERE THE SAVINGS CAME FROM ledger below the charts.
• Keep the layout readable at compact, normal, and wide widths without reverting to an undifferentiated stack of bars.

Capabilities

• Modified: human token-impact Ratatui overview and its responsive visual hierarchy.
• Unchanged: persisted telemetry, token accounting, structured CLI/MCP output, trend mode, benchmark-evidence exclusion, and the bounded wide-layout atlas.

Architecture Diagrams

• Token telemetry and dashboard ownership

No architecture change is expected; this is a visual-composition improvement within the existing TUI renderer.

Release Scope

Post-v0.4.0 backlog improvement. Do not add this issue to the v0.4.0-00 milestone and do not delay the v0.4.0 release for it.

Affected surface: token telemetry / terminal UI.

Visual References

Current composition — file reads have a full structured row, while folder walks and candidate files are cramped beneath it:

Preferred visual language — clear columns, spacing, boundaries, and label/value/bar hierarchy:

Earlier full dashboard — compact top summary, savings-composition/signal panels, and the exact source ledger as one coherent composition:

Reclaimed panel area — the removed benchmark-evidence panel demonstrates how much vertical space is now available for a clearer navigation-impact visualization:

The preferred screenshot is a style and hierarchy reference, not a requirement to combine different metrics into one row. Apply that clarity consistently to each category.

Candidate design directions to compare with real renders:

1. Give each category its own peer row/card using the preferred screenshot's label/value/bar hierarchy.
2. Keep a compact category summary near the top and use the reclaimed benchmark-panel area for a dedicated chart section.
3. Evaluate grouped vertical bars for activity share versus token-impact share across the three sources.
4. Evaluate a compact donut/pie treatment only if it keeps the remainder/other share explicit and makes the two different denominators clearer than bars.

The implementation should select the clearest bounded composition after comparing compact, normal, and wide real renders; the issue does not prescribe a chart merely for decoration.

Acceptance Criteria

• File reads, folder walks, and candidate files are immediately recognizable as three peer categories.
• Each category has a consistent visual structure and enough separation that its labels, values, descriptions, and bars cannot be confused with a neighboring category.
• File reads retain the exact observed/modelled split and proportional bars.
• Folder walks retain exact activity and token-impact values, percentages, denominators, and independently proportional bars.
• Candidate files retain exact activity and token-impact values, percentages, denominators, and independently proportional bars.
• The source ledger below the charts reconciles exactly with all three category sections.
• The layout remains bounded and readable at 80, 140, and 200 columns in dark, light, and terminal-background themes.
• Compact mode does not clip category identities or merge adjacent bar labels.
• The reclaimed benchmark-panel space is deliberately used or deliberately left as whitespace; the accepted design explains why its chart form is clearer than the alternatives.
• Any vertical-bar or donut/pie design labels both the activity and token-impact denominators, shows remainder/other share honestly, and never turns each source into an isolated 100% visual.
• Deterministic buffer tests protect arithmetic, section order, proportional bar lengths, bounds, and benchmark-attached/live equality.
• Real terminal renders are visually inspected at the supported width/theme matrix before acceptance.

Non-Goals

• Do not change telemetry collection, persisted values, accounting formulas, denominators, confidence semantics, or structured CLI/MCP schemas.
• Do not reintroduce the requested benchmark-evidence panel.
• Do not redesign the mini atlas or trend dashboard as part of this issue.
• Do not add a dependency, second dashboard framework, or interactive state solely for this composition change.
• Do not reopen or delay v0.4.0 release readiness.

Pre-Mortem

Likely failure modes:

• Repeating the rich row three times consumes too much vertical space at 80 columns.
• Stronger visual consistency accidentally makes different denominators look equivalent.
• A stylish chart consumes the reclaimed space but communicates less than the current exact bars.
• Normal-width spacing improves while compact or terminal-background themes regress.

Mitigations when this backlog item is scheduled:

• Define explicit compact, normal, and wide compositions before implementation.
• Keep each bar's numerator, denominator, and semantic label visible without relying on color.
• Compare peer rows, grouped vertical bars, and a bounded donut/pie option against the same exact data before choosing the chart form.
• Protect the complete width/theme matrix with buffer assertions and real-render review.

OpenSpec Tasks

OpenSpec change: to be assigned when this backlog improvement is scheduled.

No implementation task is created or started by this issue filing.

Privacy Check

The four reference screenshots contain only ProjectAtlas terminal UI output; no secret, token, private repository name, email address, or private filesystem path is shown.

[styler-ai]

Roadmap disposition: keep for the v0.7.0 human architecture/TUI wave, after the RC2 viewport fix. Preserve the existing accounting authority and evaluate the richer composition only against real compact, normal, and wide renders.