[kellymichels]

Two fixes to output tracking, one commit each so either can be reverted alone. Both commits pass the suite on their own.

1. Token estimates were about 44% low (12f286e)

Every script's ~N tokens est. footer, token-count.ps1 and the ztokens records divided output characters by 3.5. That was a prose rule of thumb that was never measured. Anthropic's count_tokens on 8.9M characters of real z-script output gives:

Model	chars/token
Claude Opus 5.5 / Fable 5.1 / Sonnet 5.5 / Haiku 5.5 (one tokenizer)	1.97
short, non-deploy script output alone	2.03
the previous tokenizer	2.62

• ZHelpers.ps1: one named constant, $global:ZCharsPerToken = 2.0. Each record carries its basis, est. chars/2.0.
• token-count.ps1 reads that constant instead of keeping its own 3.5.
• TOKEN_SAVINGS.md (and .txt): every measured figure is recomputed from the same captured character counts (×1.75). An active dev day is ~46,300 tokens, not ~26,500; a cached zdeploy is ~1,420, not ~810. The measurement note is rewritten. Prices are now the current models: Haiku 5.5, Sonnet 5.5, Opus 5.5, Fable 5.1. The raw-orchestration estimates were judged in tokens, so they are unchanged, apart from the full-rebuild floor, which can't be below the output it contains.
• tests/TokenEstimate.Tests.ps1 (new): checks the ratio stays in the measured band and the footer and record agree. It fails at 3.5.

2. An error or Ctrl+C left the transcript recording the window (d9cded9)

Tracking ended only on paths that reached Stop-ZTracking. After a terminating error or Ctrl+C, the transcript in %TEMP% kept capturing every later command and its output in that PowerShell window. It was never deleted, and the run was never recorded.

• All 16 tracked scripts run their body in try { … } finally { Stop-ZTracking -IfActive }, which runs on an error, an exit and Ctrl+C. The body is not re-indented, so each script's diff is three lines.
• -IfActive is silent when the script already stopped tracking itself, so the footer prints once.
• Start-ZTracking first clears leftovers (Clear-ZTrackLeftovers). It handles this window's own unstopped transcript, which is held open and named by the Process ID: in its header, and files whose process has ended, after an hour's grace. A live run's file is held open, can't be deleted, and is skipped. It never stops another window's run, or a transcript started by hand.
• tests/TranscriptLeak.Tests.ps1 (new) contains:
  • a parse-tree check that every script calling Start-ZTracking wraps the rest in that finally;
  • throw, exit and already-stopped runs in child processes with an isolated TMP;
  • both kinds of leftover, including a file held open by another process.

Tested

• Invoke-Pester .\tests with TMP isolated: 305 passed, 0 failed (1 skipped). The commit-1 state alone passes too: 296 passed, 0 failed.
• Mutation check, on the private copy: removing the try from one script, the cleanup call, the -IfActive guard or the hour's grace each fails exactly its test.
• CHECKSUMS.txt is refreshed in each commit (zchecksums -Update), and the plain-text twins are regenerated with scripts/plaintext_twins.py. --check reports them in sync.
• The diff and the new tests were scanned for private paths, hosts and project names.

The release bump comes after merge, as usual.

🤖 Generated with Claude Code