Reviewed exact bcf042f9f63f7d6e46ebf31c63d3605b07313c56, requested by Miguel. No blocking findings.

packages/cli/src/commands/init.ts:172–205 prefers display-matrix rotation over the legacy rotate tag, swaps upright dimensions for quarter turns, and selects portrait/square only for from-file. Both scaffolding callers use explicit resolution first (:964,1194); landscape and other example designs retain their previous defaults. The finishScaffold extraction preserves telemetry/transcript order.

skills/talking-head-recut/SKILL.md:266–274 expressly gives an already named ratio/platform precedence over source inference. The question now concerns layout/style/card count in both channels; it does not ask users to repeat a specified ratio. The metadata command retains rotation side data. CLI square footage uses1080×1080; the skill intentionally retains its existing near-square→4:5 preset mapping rather than promising arbitrary native dimensions.

Validation: read full init implementation/tests, from-file template and changed skill decision sections; no existing review or inline findings. My14 focused controls passed, including actual ffprobe on both committed fixtures, quarter/half turns, legacy rotation, square/odd/landscape shapes and scoped preset selection. A fresh independent pass passed21 controls, including generated real display matrices and actual rotated non-web MPEG-4→upright H.264 transcoding. Isolated caller/precedence checks are source controls, not an interactive CLI session. No full39-test rerun, TTY wizard, browser render or model-following-skill run claimed; author suite evidence remains attributed.

Nit: skills/talking-head-recut/SKILL.md:362,440 says the canvas matches/comes from the source even when an explicit user override wins or square footage maps to4:5. Say selected ratio there, and describe the inference as choosing a preset. Add a small policy pin for explicit-user-ratio precedence when maintaining the prompt tests.

Need: yes · Size: focused · Reuse: existing ffprobe/scaffold/preset path; no engine dependency introduced · Simpler: small orientation selector and shared existing post-scaffold work.

— hf-review

Verdict: APPROVE
Reasoning: Upright clip dimensions reach both init paths without overriding explicit choices, and the skill preserves user intent while removing the redundant ratio question.

Re-reviewed at 564aae14cc9c732b85c7f73e25c2d14bad8a71f0, including the delta from my bcf042f9 approval. The shared canvas writer applies the selected dimensions to stage attributes/CSS/viewport, and both init branches retain explicit-resolution precedence. All four committed video fixtures select their expected frame in independent exact-source controls.

P2 blocker — use displayed aspect ratio, including non-square pixels. packages/cli/src/commands/init.ts:172–184 derives shape solely from coded width/height plus rotation; :207–214 uses that ratio to select the canvas. A real H.264 clip encoded 720×576, SAR 64:45 has DAR 16:9. ffprobe confirms that metadata, and real Chrome reports intrinsic 1024×576. Yet the exact current probeVideo + clipFrame select 1080×1080. Before this change the template stayed 1920×1080, matching this clip; the new square frame with the existing object-fit: cover loses 43.75% of its horizontal picture. Include sample/display aspect ratio before applying rotation and choosing the nearest preset, and add an anamorphic regression fixture.

Important — qualify the no-crop claim or preserve the whole picture. templates/from-file/index.html:49 still uses cover. Even the new 48×64 (3:4) fixture cropped into a 4:5 canvas loses 6.25% vertically. Snapping improves the default but does not establish the body's “nothing is cut”; a whole-picture promise needs contain/letterboxing for unmatched ratios.

Validation: independent real ffmpeg/ffprobe fixture plus exact-source selection and browser intrinsic-dimension probe; the four committed fixture selections pass. Full init suite could not collect locally because borrowed dependencies lack @hyperframes/parsers/slideshow; no full CLI wizard/render or model-following-skill run claimed. Re-read changed init/scaffold/preset paths and tests; the unchanged skill policy/nit from my prior review remains.

Need: yes · Size: focused · Reuse: extend the existing ffprobe metadata path and canvas writer · Simpler: compute one displayed aspect ratio before preset selection.

— hf-review

Verdict: REQUEST CHANGES
Reasoning: Rotation handling and 4:5 selection work for square-pixel fixtures, but the new inference introduces a reproducible wrong-canvas regression for anamorphic video.

Re-reviewed at 8ad23620f52e12710c9ac4f94e5768975ec52d1f, including all changes since my request for changes at 564aae14.

Both concerns are resolved. packages/cli/src/commands/init.ts:173–190 applies sample aspect ratio before rotation; :200–210 now keeps arbitrary displayed ratios rather than snapping 3:4 or 4:3 to a mismatched preset. Both scaffold branches preserve explicit-resolution precedence, and explicit non-from-file designs keep their dimensions. The shared canvas writer updates the existing stage/CSS/viewport surfaces. The skill retains user-ratio precedence and its corrected probe fields.

Independent exact-source controls using real ffprobe pass all six committed fixtures, the original 720×576 SAR64:45 counterexample, combined SAR+90° rotation, 21:9 and even-size rounding/fallback cases. I reran that probe successfully. No full CLI/TTY/browser-render or model-following-skill acceptance is claimed.

Need: yes · Size: focused · Reuse: existing probe, scaffold and canvas writer · Simpler: one displayed-size calculation, shared by both init paths.

— hf-review

Verdict: APPROVE
Reasoning: The anamorphic regression and preset-snapping crop are repaired at this head, with no new blocker found in the reviewed paths.