Reader: preserve user-turn block sizes to unlock per-tool-call cost attribution

[willwashburn]

Problem

packages/reader/src/claude.ts:184-198 (captureSubagentFromToolResult) reads user turns only to index subagent markers, then discards the content. To attribute spend back to individual tool calls we need the size of each tool_result that entered context between assistant turn N and N+1.

Why this unlocks everything else

Anthropic's API reports usage at the message level — never per tool_use. But per-tool-call cost is recoverable as a delta from the session JSONL:

context(N+1) input-side tokens
  = output(N) tokens
  + sum(tool_results added between N and N+1)
  + any user free-text

Every assistant turn already has exact usage.{input, output, cacheRead, cacheCreate5m, cacheCreate1h} (captured at claude.ts:200-211). The missing piece is the size of each block the user turn contributed. Without it we cannot allocate the delta across the tool calls that caused it.

No other usage tracker does this (confirmed against TokenTracker, ccusage) — they all stop at message-level usage. This is burn's unique opening.

Plan

• Replace captureSubagentFromToolResult with a captureUserTurn that records per-user-turn block info: { toolUseId?, kind: 'tool_result' | 'text', approxTokens, byteLen, isError? }.
• Token estimate via a cheap heuristic first (bytes/4 works for ASCII-heavy content); upgrade to @dqbd/tiktoken cl100k if accuracy matters. Measure before adding a runtime dep.
• For object content, JSON.stringify before measuring.
• Extend TurnRecord with an optional sibling record or a priorUserTurn field on each assistant turn — pick whichever keeps the append-only JSONL ledger clean. Leaning toward a separate UserTurnRecord so assistant turns stay immutable once written.
• Preserve order. The reader already maintains turn order via order: string[]; user turns need to slot between assistant turns by parentUuid chain.

Acceptance

• Reader emits ordered per-user-turn block-size info alongside TurnRecord[].
• On a real session, delta reconciles: context(N+1) - cacheRead(N+1) within ±5% of output(N) + sum(new user-turn block sizes). Add a reconciliation assert as a test helper.
• Test coverage in packages/reader/src/claude.test.ts with at least one multi-turn session that exercises tool_result of varying sizes (tiny Bash output, large Read result, error results).
• No change to how existing TurnRecord consumers read data — additive only.

Depends on

Nothing. This is the prerequisite for #2.

[willwashburn]

Follow-up from investigating lazyagent: for new sessions after the hook-based ingest path lands (filed separately), the hook payload's PostToolUse event carries tool_response content directly, so per-tool-call sizing doesn't require delta math.

This issue is still needed for two reasons:

1. Historical sessions — everything spawned before hooks were installed. The delta-math path is the only way to attribute those retroactively.
2. Standalone sessions — any claude invocation not routed through relay / workforce / burn claude. We don't control all spawns in every environment.

Plan: keep this issue scoped as-is. The hook-based path (new issue) reads tool_response directly and writes TurnRecords with precise block sizes; this issue provides the same record shape from JSONL deltas as a fallback. Downstream consumers (#3 burn waste) don't need to know which path produced the data.

[willwashburn]

Updated plan after #33 (content sidecar).

With the content sidecar enabled (default), this issue's role shifts from primary to fallback. The sidecar captures the exact tool_result content verbatim, so per-tool-call token sizing can be computed by direct tokenization — no delta reconciliation needed.

Revised path ranking:

1. Primary (content.store=full, default): tokenize each tool_result in the sidecar. Exact sizes, no estimation.
2. Fallback (content.store=hash-only or sidecar missing for a historical session): the delta-math approach described in this issue's original spec. Still needed — content sidecar only captures sessions ingested after Design: content sidecar store with retention and opt-out #33 lands.

Acceptance addition: when the content sidecar is available for a turn, the reader uses it; when not, falls back to delta math with the same output shape. Consumers (#3) don't need to know which path produced the number.

Implementation impact: the reader change in this issue (preserve user-turn block sizes) is still needed, because (a) historical sessions won't have sidecar content and (b) hash-only mode explicitly skips it.