[straff2002]

Plan KL P5: the device re-runs, written up as owed with exact steps. Docs only, and nothing was run on a device.

P0–P4 are merged (PRs 937, 948, 949, 950, 951). P5 needs a phone, the glasses and two accounts, so it can't be done headlessly. This PR turns the P5 section into four runs on build 623 or later, each with what to expect and which log lines to check:

1. Signed out of ChatGPT (P0): the sign-in line once, the status-card row, and the photo line naming the sign-in; then sign in and ChatGPT answers.
2. Free-tier Groq (P1, P2): at most one 413 with limit=/requested=, then a fitted retry that keeps the tools or the allowance line, and never the Small-context advice. The repeat question is fitted before sending. This run also settles whether Groq counts the output cap.
3. A refused ElevenLabs key (P3): one 401, then no further requests, the notice once and the Settings row; the right key clears it.
4. First voice turn on a fresh install (P4): source=disclosure characters=94, no gap before turnClassified, and the introduction heard in full.

The plan's Status line and the README index row now say nothing is left to build and P5 is owed (🚧). The privacy-logging gate and gitleaks both pass.

🤖 Generated with Claude Code