Checkpoint craft: five stops, visual identity, and the direction stop first

[leochenpm]

Batch 6/8 of the release_1.6.5 sync (checkpoint craft, prompt-only). Stacked on #18. Re-mapped by hand from the net end-state of the upstream Aug 1-10 gate/orchestration iteration (c4ba98815, 707a8bf1f, 734cfc675, ed5d22684, 2182ad89a, 72e541a87).

What this adds (all prompt-level, all absent from OSS today)

1. "Show the work; stop only five times." OSS's checkpoint protocol says when to stop but never the anti-over-stopping half — that showing an artifact is not itself an ending. The five stops are now a closed set.
2. The visual preview stops once per visual identity. Narration/audio-only changes preserve the accepted silent frames; only a change to visible copy/layout/assets/windows/motion re-gates. Without this an OSS agent re-previews after every narration tweak. (Consistent with the PR5 resolver's amendment-aftermath reasons.)
3. The direction stop comes first, before any plan file exists (orchestration Gate A + video-router). OSS's compose runbook currently starts after Gate B, and nothing stops an agent from authoring the whole plan before the user has picked a direction — upstream measured 11 minutes of thrown-away authoring. An exact brief still stops, with one concept: a misunderstanding costs one message instead of a whole plan.
4. Turn mechanics that survive summarized transcripts: artifacts ride the message that ENDS the turn (mid-turn frames were never seen); an enumerated-option reply ("2", its wording, or a paraphrase) IS the decision — act on it, don't re-ask.
5. Presentation copy: progress prose is user-facing; no finding codes/severity words/internal frame roles; a passing preview message = sheet → one sentence → one question; passing checks stay silent.
6. Repair economics: all independent blocking findings from one QA result get repaired in ONE message (upstream: one-per-message cost ~10 minutes for six findings); a successful write is authoritative — never re-read to confirm it.
7. AUTO stop economics: an assembled production is ONE video — stop count fixed by the gate table, one whole-video review, an edit invalidates only its own segment.
8. gate-control: Gate B restates settled direction facts instead of reopening them; a different-language reply is a revise instruction; a recovery is executed, not narrated.

Verification

Prompt-only; the skills content test pins each new clause (stop only five times, once per visual identity, before any plan file exists, Routing ends at the direction stop, …). Full suite 266 pass / 11 skip.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VL7mRpmEphLAbcH7YmABnV