Skip to content

Fix playback, outcome-led pilot decisions and release integrity - #55

Merged
tomqwu merged 1 commit into
mainfrom
codex/review-fixes-v1-23
Sep 9, 2026
Merged

tomqwu merged 1 commit into
mainfrom
codex/review-fixes-v1-23

Conversation

@tomqwu

@tomqwu tomqwu commented Sep 9, 2026

Copy link
Copy Markdown
Owner

This resolves the eight v1.22.0 review findings. Presenters keep their Auto-next choice, lifecycle failure exploration pauses conflicting narration, and delayed media events cannot reclaim the diagram after manual takeover.

Pilot continuation now follows a sponsor-approved objective instead of requiring every quality or reliability case to release effort or cash. The banking case gains a three-state observation worksheet that separates modernization from incremental AI effort, keeps missing or mismatched controls unknown, retains review/rework/setup/operating costs, and exposes slowdowns. Ranking-only discovery answers no longer become percentages. Shared-platform readiness follows the actual workflow being reused.

Publication checks now bind every narrated destination to its reviewed rendered content and reject conflicting release/tag identities before Pages deployment. One pilot-decision recording was refreshed with measured captions; the other affected explanations were explicitly reviewed and retained. All four PDFs advance to v1.23.0, retaining 116 slides and existing anchors.

Validation:

  • Chromium and WebKit: Auto-next controls and real audio completion; lifecycle caption anchors, reverse seek, failure branches, replay, keyboard takeover and reduced motion.
  • Both browsers: workflow-specific readiness, matched three-state arithmetic, missing controls, worksheet download, mobile layout and no-JavaScript form submission.
  • Existing cross-tab/nested-player audio-focus suite and 139 caption-aligned diagram states pass in both browsers.
  • Model and release tests pass, including modernization helping while AI slows work, an evidenced quality case with zero cash capture, missing cash evidence, conflicting tags, annotated tags and safe retries.
  • All 116 recordings/captions agree with scripts; 167 narration destinations have current content reviews.
  • All four PDF exports rendered and reviewed; 116 PDF pages, diagram labels, edition markers, links, scope and research provenance verified. Full repository CI runs before merge and deployment.

Closes #47
Closes #48
Closes #49
Closes #50
Closes #51
Closes #52
Closes #53
Closes #54

@tomqwu
tomqwu merged commit f169fc7 into main Sep 9, 2026
3 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment