Chantier: close the seven-requirement capability gaps

[LeadcodeDev]

Context

A capability audit measured this engine against seven requirements for producing video in the register of two reference films (Machina by 1600.agency, Aikido). Four requirements were previously audited and implemented; they hold. Three were new, and two of them exposed the most serious obstacles found so far.

Evidence base: 450 fixtures, 526 rendered frames decoded and diffed at the pixel level, a measured typographic sweep, and a root-cause diagnosis proven by four falsifiable predictions verified in advance.

What holds. The hardest requirement — no perceptible scene boundaries — is objectively met: a standard shot-change detector finds zero cuts in the template, matching Machina (0) and Aikido (1). Camera work and the word-by-word reveal are sound.

What blocks. Every graphic element belongs in a card in this style, and 23 of 57 components render zero pixels inside one without an explicit height. The typographic register needs tight negative tracking, and the line-breaking pass ignores letter-spacing entirely. Background crossfades step by a full LSB across the frame at every junction.

Workstreams

Each is an independent PR off main with a disjoint file set, so they can land in any order.

• W1 · Transitions — float crossfade, static hard cut, transitions never reaching their final state
• W2 · Typography — letter-spacing ignored when breaking lines; relative units resolving to zero
• W3 · Component sizing — the 23 components with no size source
• W4 · Painters that leave their box — stat, badge, kbd, caption, timeline
• W5 · Geometry validation — content checked only for text, never against its card, rotation ignored, announced duration wrong
• W6 · PNG sequence ordering — frames permuted by a shared atomic
• W7 · Documentation — the typography rule forbids the target register; the card rules are measurably backwards

Verification, applying to every workstream

cargo test --workspace && cargo clippy --workspace --all-targets
for f in examples/*.json; do ./target/release/rustmotion validate -f "$f"; done

Every visual claim must be proven by rendering and measuring pixels. A passing unit test is not evidence that something is visible on screen. Use ./target/release/rustmotion — never the binary on $PATH, which is months stale.

If a change makes an existing example fail, report it rather than weakening the check. That signal is wanted.