Honor fidelity in burn compare (#95)

CHANGELOG.md

@@ -6,6 +6,10 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/)
 
 ## [Unreleased]
 
+### Changed
+
+- **`burn compare` honors fidelity** ([#95](https://github.com/AgentWorkforce/burn/issues/95)). The aggregate now defaults to the `usage-only` fidelity floor — turns whose per-turn token coverage is `aggregate-only` / `cost-only` / `partial` are excluded so a session with mixed fidelity can't silently bias the cost/turn or one-shot rate of full-fidelity peers from the same model. New `--fidelity <class>` and `--include-partial` flags override or disable the floor; coverage notes and the `--json` output gain a `fidelity` block (`{ minimum, excluded, summary }`) computed against the unfiltered slice. Records emitted before `TurnRecord.fidelity` existed still pass for backward compatibility.
+
 ## [0.33.0] - 2026-04-27
 
 ### Added

README.md

@@ -268,7 +268,7 @@ You can override per-call via `costForUsage(usage, model, pricing, { reasoningMo
 burn summary [--since 7d] [--project <path>] [--session <id>] [--workflow <id>] [--agent <id>] [--provider <p>] [--by-provider]
 burn by-tool [--since 7d] [--project <path>] [--session <id>] [--provider <p>]
 burn waste [--since 7d] [--project <path>] [--session <id>] [--workflow <id>] [--provider <p>]
-burn compare [--models a,b] [--since 7d] [--project <path>] [--session <id>] [--workflow <id>] [--agent <id>] [--min-sample <n>] [--json|--csv]
+burn compare [--models a,b] [--since 7d] [--project <path>] [--session <id>] [--workflow <id>] [--agent <id>] [--min-sample <n>] [--fidelity <class>] [--include-partial] [--json|--csv]
 burn claude  [--tag k=v ...] [-- <claude args>]
 burn codex   [--tag k=v ...] [-- <codex args>]
 burn opencode [--tag k=v ...] [-- <opencode args>]
@@ -298,7 +298,9 @@ This is observed data, not counterfactual: it tells you what happened when you a
 
 Standard filters apply: `--session <id>` limits to a single session, `--agent <id>` limits to a stamped agent ID, `--workflow <id>` to a stamped workflow ID, `--project <path>` to a project path or git-canonical projectKey.
 
-Output formats: TTY table (default), `--json` for scripts, `--csv` for spreadsheets. `--json` and `--csv` are mutually exclusive.
+By default, `burn compare` only aggregates turns with `usage-only` fidelity or better — `aggregate-only`, `cost-only`, and `partial` turns are excluded so a session with mixed fidelity can't silently bias the cost/turn or one-shot rate of full-fidelity peers from the same model. When the gate dropped anything, the table prints an `excluded N turns below <class> fidelity (… aggregate-only, … cost-only, … partial)` coverage note. Override the floor with `--fidelity full | usage-only | aggregate-only | cost-only | partial`; `--include-partial` is shorthand for `--fidelity partial` and includes every turn. Records emitted before `TurnRecord.fidelity` existed always pass for backward compatibility.
+
+Output formats: TTY table (default), `--json` for scripts, `--csv` for spreadsheets. `--json` and `--csv` are mutually exclusive. The `--json` payload includes a `fidelity` block (`{ minimum, excluded, summary }`) computed against the unfiltered slice so consumers can render their own coverage UI.
 
 ### `burn rebuild --reclassify` — backfill activity labels on old turns
 

packages/analyze/CHANGELOG.md

@@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Changed
+
+- **`hasMinimumFidelity` and `summarizeFidelity` are now wired into `burn compare`** ([#95](https://github.com/AgentWorkforce/burn/issues/95)). No API change in `@relayburn/analyze` itself — this entry just records the consumer-side adoption of the helpers shipped in 0.14.0 ([#41](https://github.com/AgentWorkforce/burn/issues/41)). See `@relayburn/cli` for the CLI surface (`--fidelity`, `--include-partial`, the new JSON `fidelity` block, and the "excluded N turns" coverage note).
+
 ## [0.33.0] - 2026-04-27
 
 ### Added

packages/cli/CHANGELOG.md

@@ -11,6 +11,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - **`burn limits` honors fidelity on its 5-hour forecast** ([#105](https://github.com/AgentWorkforce/burn/issues/105)). The forecast still consumes every windowed turn — partial / aggregate-only / cost-only data still contributes to the running token total — but `burn limits` now classifies the contributing slice via `summarizeFidelity` and surfaces a binary `high` / `low` confidence flag. Text mode appends a `forecast: low-confidence (N of M contributing turns lack per-turn token data)` notice when at least one contributing turn is missing per-turn token coverage; full-fidelity windows print no notice. `--json` output gains a `forecast.fidelity` block carrying the `confidence` flag and the underlying `FidelitySummary`. `--watch` re-evaluates confidence on each tick so the flag flips as fresher full-fidelity turns land.
 
+### Changed
+
+- **`burn compare` honors fidelity** ([#95](https://github.com/AgentWorkforce/burn/issues/95)). The aggregate now defaults to the `usage-only` floor: turns whose fidelity is `aggregate-only`, `cost-only`, or `partial` are excluded so a session with mixed fidelity can't silently bias the cost/turn or one-shot rate of full-fidelity peers from the same model. Records emitted before `TurnRecord.fidelity` existed (pre-#41 ledgers) still pass for backward compatibility. New flags: `--fidelity <class>` (any of `full | usage-only | aggregate-only | cost-only | partial`) overrides the floor; `--include-partial` is shorthand for `--fidelity partial` and includes every turn — both invalid combinations exit 2 with a clear message. Coverage notes gain an `excluded N turns below <class> fidelity (… aggregate-only, … cost-only, … partial)` line whenever the gate dropped anything, the JSON output gains a top-level `fidelity` block (`{ minimum, excluded, summary }`) computed against the unfiltered slice, and per-model totals render `—` instead of `$0.00` when a model survived the filter with zero turns. When fidelity filtering is active (the default) `burn compare` falls back to the in-memory `queryAll` path so the gate is correctly applied; `--include-partial` (or `--fidelity partial`) reuses the archive's grouped SQL path from #88.
+
 ## [0.33.0] - 2026-04-27
 
 ### Added