Honor fidelity in burn limits (#105)

[willwashburn]

Summary

• burn limits consumes the entire windowed slice (no refusal of partial / aggregate-only / cost-only data) and now classifies the contributing turns via summarizeFidelity to derive a binary high / low confidence flag for the 5-hour forecast.
• Text mode appends forecast: low-confidence (N of M contributing turns lack per-turn token data) to the forecast block when any contributing turn lacks per-turn token coverage; full-fidelity windows print no notice.
• --json gains a forecast.fidelity block: { confidence: 'high' | 'low', summary: FidelitySummary }. --watch re-evaluates the flag on each tick.

Test plan

• pnpm run build
• pnpm run test:ts (518 tests pass)
• New packages/cli/src/commands/limits.test.ts cases:
  • high-confidence (all full turns) renders no notice
  • low-confidence (one partial turn) appends the notice but still renders the burn rate + projection
  • --json carries forecast.fidelity.{confidence, summary} with the right counts
  • --json reports confidence: 'high' when every turn is full
  • --watch-style re-invocation flips low → high as new full-fidelity turns arrive

Refs

• Closes burn limits: honor fidelity (mark forecasts low-confidence on partial usage) #105
• Refs Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 (original fidelity scope), Add coverage and fidelity metadata to TurnRecord (#41) #76 (helper landing PR)

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 3 additional findings in Devin Review.

[devin-ai-integration]

🟡 Interface comment formula for lowConfidenceTurns incorrectly excludes unknowns, contradicting the implementation

The doc comment on lowConfidenceTurns at line 56 claims the value is total - (full + qualified usage-only) - (any unknowns), which says unknown/no-fidelity turns are subtracted from the count. However, the implementation at limits.ts:579-581 does the opposite: turns with !f (no fidelity) are added to lowConfidenceTurns. The function-level doc at limits.ts:569-571 correctly states unknowns "get counted toward lowConfidenceTurns". The actual formula is total - full - qualified_usage_only (unknowns are included, not excluded). This is an exported interface (ForecastFidelity), so downstream consumers of the type definition will get an incorrect mental model of what the field contains — e.g. for a window with 5 turns (2 full, 1 partial, 2 unknown), the comment says lowConfidenceTurns = 5 - 2 - 0 - 2 = 1 but the actual value is 3.

Suggested change

	// Count of turns whose per-turn token data is unreliable for forecasting.
	// Equivalent to `total - (full + qualified usage-only) - (any unknowns)`;
	// surfaced separately so the rendered notice can read "N of M".
	lowConfidenceTurns: number;
	// Count of turns whose per-turn token data is unreliable for forecasting.
	// Equivalent to `total - (full + qualified usage-only)` — unknowns (records
	// with no `fidelity` field) are counted here too, not excluded.
	// Surfaced separately so the rendered notice can read "N of M".
	lowConfidenceTurns: number;

---

Was this helpful? React with 👍 or 👎 to provide feedback.