[miguel-heygen]

What was wrong

A clip bound with data-var-src inside a sub-composition rendered frozen at its authored placeholder's length when the mount set the real file through data-variable-values. Reported on 0.8.206 (rated 4/10): a <video data-start="0" data-duration="60" data-var-src="src"> over a 3 s placeholder, mount value a 13 s clip; the render played the first 3.1 s and held the last frame. No error anywhere.

The compiler can only read the authored src, so media bound to a variable is sent through a browser probe that reads what the runtime resolved. hasVariableBoundMedia, the gate for that probe, only asked whether the render's own --variables named the element's variable. A value from the mount's data-variable-values, or a default the composition declares, never opened it, so frame extraction and the audio mix used the placeholder.

Fix

hasVariableBoundMedia asks the question the runtime answers: does the element's variable have a value anywhere it looks: the render's variables, the data-variable-values of a mount above the element, or a declared default (data-composition-variables)? It reads them with parseHostVariableValues and readDeclaredDefaults, the readers the bundler and runtime already use. A variable with no value anywhere still skips the browser, as before.

Design calls:

• Fixed at the gate, not by resolving the source in the compiler: the browser probe already reconciles source, duration and audio for the render-variable case, so the mount and default cases go through the same path.
• A declared default counts even when it equals the placeholder: a few more renders launch the probe, and none can be wrong.

What I measured

fails without the fix: against main's gate, the new real-render test fails (expected 0 to be greater than 200: the frame at 6 s is the black placeholder, not the white part of the bound clip), and the three new gate and compile cases fail (70 pass, 3 fail); with the fix all pass (73 pass in probeStage plus the compile fixture, miga, Linux).

The real render (variableMountMedia.integration.test.ts): a 3 s black placeholder, a 13 s clip that is black for 5 s then white, a mount binding the 13 s clip, rendered for 8 s; the frame at 1 s is black and the frame at 6 s is white. The compile fixture (htmlCompiler.variableMedia.test.ts) uses the same 3 s and 13 s files. Both are added to the integration lane list. Fallow audit, oxfmt, oxlint and the reachability check pass locally.

What I did NOT exercise

A default a sub-composition declares on its own file's <html data-composition-variables>: inlining drops that attribute, so the gate cannot see it and a clip bound to such a default (with no mount value and no render variable) still uses its authored src. Only a default declared on the root <html> or a composition element is read. The mount-value case that was reported is covered.

The Studio preview: it resolves bindings in the browser runtime (applyVariableBindings), which this does not touch, and I did not open it. The audio side of the same bound clip: the probe reconciles audio entries by the same path, but the test measures picture only. A variable bound on an <img> or <source>, and a declared default that differs from the placeholder, are covered by the gate's unit cases, not by a render.

[github-actions]

Edit accuracy: accurate 271 (base branch 271), smooth 126 of those

The gate passes.
Smoothness is reported in the artifact, not gated. A case fails only if it fails 2 of 3 runs.

Quarantined, measured but not gated (0)

Unstable (1)

• seqplainkeys-none-px-r0-root-z100: tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo false, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true / tracking 0.06, pressJump 0, drop 0, reload 0, render 0.03, renderKey -, undo true, teleport true

[heygengenesis]

Verdict: REQUEST CHANGES · 1 blocker · 0 important · 0 nits

The fix itself looks right. The blocker is the required Comments check, which fails on this head (inline).

Scope

Read probeStage.ts, its tests, the compile fixture, the real-render test and the integration lane list at 386b4016, plus parseHostVariableValues/readDeclaredDefaults (both tolerate malformed JSON) and the loader's mount scope (stashInstanceVariables: declared defaults plus host values). Locally: bun test on htmlCompiler.variableMedia and probeStage passes 73/73. CI is green except Comments; Windows render was still running.

Not a finding, unverified: a sub-composition that declares its default on its file's <html data-composition-variables> (the loader reads doc.documentElement) loses that attribute in the compiled HTML. A scratch compile shows hasVariableBoundMedia returning false there, while the template-root form returns true. I did not check whether the render runtime still resolves that default.

[heygengenesis]

Verdict: APPROVE — 1 resolved, 0 open, 0 new · 386b401 → c48378e. The change is comment-only, and CI is green including Comments. The PR body now names the <html>-declared default as not covered.