fix: MODEL DISTRIBUTION shows retired models (4.5) from stats-cache historical data

[pitimon]

Problem

MODEL DISTRIBUTION donut shows Haiku 4.5 (57%), Opus 4.5 (34%) — but these models have zero active sessions. Only Opus 4.6 and Sonnet 4.6 are currently in use.

Root Cause

Donut data comes from stats-cache.json (modelUsage) which is cumulative all-time totals from Anthropic SDK — includes every model ever used.

stats-cache (historical):          JSONL sessions (actual):
  haiku-4.5:   122M cache tokens     (no sessions)
  opus-4.5:    1.1B cache tokens     (no sessions)
  sonnet-4.5:  68M cache tokens      (no sessions)
  opus-4.6:    2.5B cache tokens     199 sessions ← active
  sonnet-4.6:  748M cache tokens     1,833 sessions ← active

The donut includes cache tokens from retired models, making them appear dominant (Haiku 4.5 at 57%) when they haven't been used in months.

Dual Source Problem

This is the same dual-source issue documented in ADR-002:

• stats-cache.json → authoritative for all-time totals (includes retired models)
• JSONL sessions → authoritative for current/windowed data (only active models)

The donut should use JSONL-derived model data (or filter stats-cache by active models).

Fix Options

Option A — Filter by active sessions (recommended)

Only show models that have at least 1 session in the selected date range:

// Cross-reference with sessions to find active models
const activeModels = new Set(sessions.map(s => s.model).filter(Boolean));
const data = Object.entries(modelUsage)
  .filter(([model]) => activeModels.has(model))
  .map(...)

Option B — Use JSONL session tokens for donut

Aggregate tokens from sessions directly instead of stats-cache:

const modelTokens = new Map();
for (const s of sessions) {
  if (!s.model) continue;
  const current = modelTokens.get(s.model) || 0;
  modelTokens.set(s.model, current + (s.input_tokens || 0) + (s.output_tokens || 0));
}

This respects the date range filter (from/to) and only shows models actually used in that window.

Option C — Show both with visual distinction

Keep all models but visually distinguish active vs retired:

● Opus 4.6 (33%)    ← solid color (active)
● Sonnet 4.6 (15%)  ← solid color (active)
○ Haiku 4.5 (25%)   ← dimmed/striped (retired)
○ Opus 4.5 (20%)    ← dimmed/striped (retired)

Data Evidence

Model usage from JSONL sessions (current):
  claude-sonnet-4-6:  1,833 sessions (Mar-Apr 2026)
  claude-opus-4-6:      199 sessions (Feb-Apr 2026)
  <synthetic>:           14 sessions (Mar 2026)

Model usage from stats-cache (all-time, includes retired):
  haiku-4.5:    122M cache tokens — NO active sessions
  opus-4.5:     1.1B cache tokens — NO active sessions
  sonnet-4.5:   68M cache tokens — NO active sessions

Context

• Same dual-source pattern as ADR-002 (JSONL for windowed, stats-cache for all-time)
• Related: fix: cache token percentage shows 38,392% — wrong denominator in calculation #56 (cache token % fix), feat: improve chart professionalism — bar chart for projects, enhanced donut for models #130 (chart improvements)

[pitimon]

Picking up Option B — compute model tokens from JSONL sessions instead of stats-cache. Only active models will show.

[pitimon]

Resolved in PR #141.

• Model donut now computes I/O tokens from JSONL sessions instead of stats-cache
• Only active models shown (Opus 4.6 + Sonnet 4.6) — retired models (Haiku 4.5, Opus 4.5) excluded
• ADR-002 pattern: JSONL for current data, stats-cache for all-time totals