Honor fidelity in burn plans (#108)

[willwashburn]

Summary

• computePlanUsage now annotates each cycle with a fidelity: { confidence, summary } block. confidence === 'high' only when every contributing turn is full or usage-only with both per-turn input and output token coverage; otherwise low. Records without a fidelity field stay best-effort high (matches the codebase's existing pre-Coverage and fidelity metadata: distinguish missing, zero, aggregate-only, and partial usage #41 backward-compat policy). Spend totals continue to include partial / aggregate-only / cost-only contributions — under-counting silently is worse than annotating low-confidence — so spentUsd becomes the lower bound the consumer renders against the new flag.
• burn plans (list view) gains a confidence column when at least one plan has any low-confidence cycle, and a footer note naming the affected plan + lower-bound caveat (note: claude-pro: 3 of 412 turns this cycle lack per-turn token data — totals are a lower bound.). Full-fidelity cycles render exactly as before — no extra column, no footer.
• --json emits a per-plan usage.fidelity: { confidence, summary } block carrying the same FidelitySummary shape summarizeFidelity produces elsewhere.
• PlanUsageFidelity is exported from @relayburn/analyze.

Rebase consideration for PR #131

PR #131 (issue #91) is migrating burn plans to read from archive.sqlite. This PR targets the current queryAll-based path on main. When #131 lands, the rebaser needs to apply the same fidelity-collection + low-confidence rule to the archive-backed path so the new annotations don't disappear. The annotation logic itself lives entirely in computePlanUsage (analyze) — it walks whatever turns the caller hands it — so the fidelity data only needs to keep flowing through to the renderer. Concretely, the conflicts will be in packages/cli/src/commands/plans.ts runList and the helper that loads turns; everything in packages/analyze/src/plan-usage.ts should rebase cleanly.

Test plan

• pnpm run build clean.
• pnpm run test:ts — 523 passing (10 new).
• New analyze tests cover: high-confidence cycle (all full), high-confidence cycle (usage-only with both axes), high-confidence cycle for unknown-fidelity (older ledger writers), low-confidence cycle (partial turn), cost-only contributions counted toward spend + flagged low-confidence, empty cycle, and out-of-cycle turns ignored.
• New CLI tests cover: text table omits the column + footer when every cycle is full-fidelity, text table renders the confidence column + footer note when any cycle is low-confidence, and --json emits the usage.fidelity block with the right confidence / summary shape.

Refs

Closes #108 — refs #41, #76 (which shipped summarizeFidelity / hasMinimumFidelity).

---

[devin-ai-integration]

Devin Review found 1 potential issue.

⚠️ 1 issue in files not directly in the diff

⚠️ Missing root CHANGELOG entry for cross-package change (AGENTS.md violation) (CHANGELOG.md:7-11)

This PR touches both packages/analyze (new PlanUsageFidelity type + deriveFidelity logic) and packages/cli (fidelity column + footer in burn plans list view + JSON). AGENTS.md states: "Update [Unreleased] only when the work spans packages or warrants a top-level summary; single-package work belongs only in that package's CHANGELOG." Since the work spans two packages, the root CHANGELOG.md's [Unreleased] section should have an entry for #108, but it does not (CHANGELOG.md:7-11).

View 3 additional findings in Devin Review.