[wasikarn]

Refs #379 (item 3 of 3; items 1 and 2 merged in #380 and #384).

What

• raw_records (subagent files) and scan_transcript (main transcript) each held a copy of the usage-record
  builder. The copies differed only in t: and f:, and a past cw1h fix had to edit both. That builder is
  now one def usage_rec($t; $f) in usage_def, next to adv_def. Both readers call it.
• emit_rows and the orchestrator path each built the jq_failed sentinel row. Now both call
  jq_failed_row <stream> <files>.
• No behaviour change. cost-tracker.sh: cache writes always priced at 5m TTL, never 1h TTL #162/cost-tracker: cache-write remainder dropped + pre-1.1.116 opus-5-5 mispricing unmarked #163 (the 5m/1h split) moved verbatim into the def. The verify-window logic
  is untouched.

Proof

Equivalence. The develop copy (with #384 merged) was frozen and both copies ran on the same inputs.
Each run had a throwaway HOME, settle 0, and a date shim. Every run compared costs.jsonl, the marker,
stdout, stderr and the exit code byte for byte. The only normalization: bash's <script path>: line N:
prefix.

corpus	bash 3.2	bash 5
316 real sessions (all 276 with subagents, 5 with codex calls, 40 without subagents)	1,531 rows, 0 diffs	1,531 rows, 0 diffs
8 hand-made sessions: both sentinel paths, 1h/multi-iteration writes, codex in main and subagents, every meta shape	61 rows, 0 diffs	61 rows, 0 diffs

Tests. Phase 6 of test-cost-tracker.sh:

• The same two-iteration line (1h writes 300 + 400, flat 1000) sits in the main transcript and in a subagent
  file. Both rows must be cw 300 / cw1h 700. No test covered the subagent path's split before.
• A usage: "not-an-object" line in each stream must give each stream the exact sentinel row (all keys)
  and its stderr line.

Red on a develop copy whose raw_records copy alone drifts (ignores iterations), which is the bug class
this removes. It passes on the develop copy and on the new code under bash 3.2 and 5. Mutants killed
(across test-cost-tracker.sh and test-session-stop.sh), 10 of 10:

• t dropped or f wrong at either call site
• the def ignores iterations
• the def's t fixed to null
• sentinel args swapped
• sentinel files emptied
• sentinel stream key dropped

Timing. Whole hook, 20 interleaved runs, medians (p90): unchanged, as expected for a refactor. At
193 subagents: 3,783 (3,968) -> 3,783 (4,044) ms on bash 3.2, 3,891 (4,248) -> 3,847 (4,059) ms on bash 5.
At 6 subagents: 155 -> 157 / 176 -> 176 ms.

Skipped: single-pass codex tally

emit_codex_invocations re-reads the subagent files that raw_records already read. On the 193-subagent
session that second read costs about 1.15 s (5 runs, 1,122-1,197 ms), about 30% of the hook. Folding it
into raw_records would merge two failure domains: today a malformed subagent line gives a sentinel row,
and the codex tally still counts. Keeping them apart needs the per-field try/catch that scan_transcript
uses. The local corpus has 0 subagent files with a codex call, so only hand-made fixtures could cover it.
Left as a follow-up rather than mixed into this refactor.