burn plans: honor fidelity (mark monthly spend totals low-confidence on partial usage)

[willwashburn]

Context

PR #76 (#41 first cut) shipped summarizeFidelity and hasMinimumFidelity in @relayburn/analyze, but stopped short of wiring those helpers into the commands that consume the ledger. Quoting PR #76's deferred-work paragraph:

burn compare, burn waste, burn limits, burn plans behavior gating on fidelity class — the helpers (hasMinimumFidelity, summarizeFidelity) are now in place; wiring them into each command is the natural follow-up.

burn plans (packages/cli/src/commands/plans.ts) lists plans and runs computePlanUsage against queryAll(...) to render monthly spend against each plan's budget. computePlanUsage sums costForTurn(...) across the cycle, so a turn whose source omitted token fields contributes \$0 and silently makes a plan look under-budget. Quoting #41:

burn limits / burn plans (#5, #39) … should permit partial usage data where enough exists for spend totals … should mark projections as low-confidence when the underlying fidelity is partial.

Proposal

In packages/cli/src/commands/plans.ts (specifically the runList path that renders plan status) and computePlanUsage (packages/analyze/src/plan-usage.ts):

1. Permissive filter. Like limits, plans is allowed to include partial / aggregate-only / cost-only data — incomplete totals are still useful as a lower bound. Do not default-exclude.
2. Annotate per-plan confidence. Extend PlanUsage with a fidelity field: { confidence: 'high' | 'low', summary: FidelitySummary } computed over the cycle's contributing turns. confidence === 'high' when every turn in the cycle is full or usage-only with hasInputTokens && hasOutputTokens; otherwise low.
3. Surface in the human-readable table. When any plan's cycle has low confidence, render a note column or a footer line: note: 3 of 412 turns this cycle lack per-turn token data — totals are a lower bound. Suppress when all cycles are full (avoid noise in the common case, matching summary's pattern).
4. JSON contract. Where plans already emits JSON (the --json path on runList), add the fidelity block per plan. Programmatic consumers reading monthly spend get the same confidence + FidelitySummary shape summary --json already establishes.
5. cost-only sources count toward spend. A cost-only source provides a price even when it provides no tokens — spend totals should include it. Mark the plan low-confidence on the token-coverage axis but do not drop the cost.

Acceptance criteria

• burn plans continues to render monthly totals when the cycle slice contains partial / aggregate-only / cost-only turns.
• Each rendered plan row shows a low-confidence note when any contributing turn lacks per-turn token coverage; full-fidelity cycles show no note.
• PlanUsage exposes a fidelity field ({ confidence, summary }) and --json emits it per plan.
• cost-only source contributions count toward spend totals, with the cycle marked low-confidence on the token-coverage axis.
• New tests in packages/analyze/src/plan-usage.test.ts cover: high-confidence cycle (all full), low-confidence cycle (partial turn), cost-only contributions counted toward spend.
• New tests in packages/cli/src/commands/plans.test.ts cover the rendered note and the JSON shape.

Out of scope

• The add / remove / set-reset-day subcommands — they don't read the ledger.
• Cursor-style providers that lack spend tracking entirely (Collector: Cursor — won't implement (data moved server-side) #22) — they already render the "spend tracking unavailable" note independent of fidelity.
• Per-day breakdowns of confidence within a cycle. Cycle-level binary flag is enough for the first cut.

Refs

• Spun out of Add coverage and fidelity metadata to TurnRecord (#41) #76 (Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 follow-up)