[G3] Accounting fixes unlocked by shared evidence: Codex per-request turns (#504), subagent project rollups and harness split (#506), `sessions list` token column

[willwashburn]

Part of #553 (group 3 — depends on burn #562 cutover; relayhistory #172 usage records, #175 project identity, #170 relationships; parallel with #560).

Problem

Two open accounting defects are cheaper to fix on top of relayhistory's evidence than in burn's readers, and the reader fix path closes at cutover:

• Codex turn accounting collapses many API requests into one turn #504 Codex turn accounting collapses many API requests into one turn (42 recorded turns vs 2,545 real requests in one session; per-turn metrics off by ~60×). PR Fix Codex request-level turn accounting (#504) #512 is open against the old reader. relayhistory Ghost surface: detect Codex/Claude slash-command invocations from session content #172 produces one session_requests row per Codex token_count delta (per API request) and cli: fold limits and plans under one burn budget verb with nested plans subcommand #164 stamps turn_id, so burn can materialise inferences at request granularity and keep turns at task granularity — both denominators become available.
• Attribution and grouping gaps: subagent spend, project rollups, harness split, session token column #506 subagent spend does not roll up to the parent project (100 orphaned subagent sessions; chief undercounted 290M vs 462M tokens); no --group-by project; no harness split in --bucket; no token column in sessions list. relayhistory Add WasteFinding/WasteAction structured envelope (closes #56) #175 gives every session a project_key with method = Inherited for delegated children, and Fold burn by-tool into burn summary --by-tool (closes #156) #170 adds continuity edges, so the rollup is a join, not a heuristic.

Scope

1. Codex requests: inferences rows built from SessionEvidence.requests (one per request; request_id_source = "codex_token_count"), turns still one per task_started/task_complete; burn compare/summary --quality per-turn metrics gain --per request|turn (default request for Codex, documented in README.md); close Codex turn accounting collapses many API requests into one turn #504 and either land the request-level parts of PR Fix Codex request-level turn accounting (#504) #512 or close it superseded.
2. Project rollups: summary, hotspots, findings gain --group-by project keyed by project_key; delegated children attribute to the parent's project_key (relayhistory method = Inherited) and --by-relationship still shows them separately; --bucket output gains a harness dimension (SourceKind).
3. sessions list: add tokens (total) and cost columns from the ledger (no relayhistory call; keep the verb offline).
4. Tests on the corpus: nested-subagent rolls up to the parent project; Codex multi-turn yields N inferences for M token_count events; golden CLI snapshots updated with a documented diff.
5. Close Codex turn accounting collapses many API requests into one turn #504 and Attribution and grouping gaps: subagent spend, project rollups, harness split, session token column #506 with links; update CHANGELOG.md.

Acceptance

• burn summary --session <codex fixture> --json reports turns and requests separately and cost_per_request uses the request count.
• burn summary --group-by project on nested-subagent shows one project row whose tokens equal parent + children.
• burn sessions list shows a tokens column; --json includes total_tokens.

Out of scope

New harness support (#560).