fix(render): make the manifest's complete flag mean something

[ringoliRob]

complete was written as a literal True on every capture, so nothing in tiles.json could ever distinguish a full page from one viewport of it. The signal that would have revealed the difference — a measured page height equal to the emulated viewport (i.e. the tile height) — was neither warned about at render time nor recorded in a form a later consumer could check, since the tile height itself never made it into URL manifests.

This is independent of any particular truncation bug: the same silence held for the height formula #131 fixed and will hold for whatever causes the next one.

• page_metrics.truncation_reason() states the rule once, next to the shared measurement JS, so the standard and turbo backends agree on what a manifest means whichever Chrome the capture ran on.
• Both backends log a warning and write complete: false when the probe fell back or the height came back at exactly one viewport. The tiles are still captured and written — only the claim about them changes.
• URL manifests now record tile_height and viewport_width, so a consumer can redo the comparison instead of being told the geometry out of band.
• The standard path now treats a zero/negative measured height as a probe failure, matching the turbo path; it used to tile nothing and report success.
• pixelbrowse's SKILL.md reads tiles.json instead of guessing: the "do not run ls, just read tile_0000.jpg" rule made a truncated capture indistinguishable from a short page, and a failed render indistinguishable from an empty one.

Note for the index pipeline: complete: false makes the embed stage skip a tile directory, so an affected page is now dropped from an index build rather than indexed as a fraction of itself.

[vercel]

@ringoliRob is attempting to deploy a commit to the andylizf's projects Team on Vercel.

A member of the Team first needs to authorize it.