fix(live-proof): only publish recordings that demonstrate a change

[steipete]

Summary

Every live proof published so far was useless, and I only caught it by watching the videos rather than checking that the URLs resolved.

• Terminal clips were 0.9s: an empty prompt, the typed command, end — before any output rendered. expect_output captured the pane exactly once with no polling immediately after Enter, so it threw instantly; it also grepped the whole pane including the echoed command, so other plans could match the command the model had just typed.
• Browser clips ran 16s with every frame identical, because plans asserted static text already rendered at load (feat(bay): focus the crustacean finder with the / shortcut #1182 asserted "press /", the finder's own hint) and nothing scrolled the changed region into frame.
• Nothing gated attachment: drive_status was recorded but unused, so failed drives were uploaded and posted to public review comments.

Fixes: terminal run waits for real output and expect_output polls with a timeout against post-command content only; both drivers settle between steps, hold the end state, and honour a minimum length; browser steps scroll their target into view.

The substance gate is semantic, not pixel-based. Attachment now requires an expectation that was absent at plan start and satisfied afterwards — a demonstration that verifies nothing publishes nothing. Pixel-difference gating was implemented, measured against real artifacts, and rejected: the junk frozen-page recording scores SSIM 0.914 between first and last frame while a genuine blade-opening demo scores 0.980, because ambient page animation swamps any threshold. The review prompt now requires a state-changing action plus an assertion on content that is absent beforehand, and real selectors rather than prose.

Validation

Run locally against the two real plans:

• the organic feat(bay): focus the crustacean finder with the / shortcut #1182 plan that produced the frozen video now logs skip: plan verified nothing that changed and writes no manifest — nothing would be attached;
• a plan asserting "Found openclaw/clawsweeper#1178" (absent at load) attaches, and its extracted frames show the input filling, the status flipping, and the toast appearing.

pnpm build clean; live-proof suite 33/33; full unit suite 2,326 passed / 0 failed (parallel with review, exit 0). Autoreview (Codex, gpt-5.6-sol, high): clean, "patch is correct (0.98)".

Follow-up

The five PRs that received junk videos need cleanup once this lands: re-dispatching the same heads overwrites the head-scoped R2 objects and record block, and any plan that legitimately has nothing to show needs its stale Live Proof block stripped from the record.

[clawsweeper]

🦞👀
ClawSweeper picked this up.

Pull request received. I will update this pull request when review starts.

[clawsweeper]

ClawSweeper status: review started.

I am starting a fresh review of this pull request: fix(live-proof): only publish recordings that demonstrate a change This is item 1/1 in the current shard. Shard 0/1.

This placeholder means the worker is alive and reading the current context. I will edit this same comment with the actual review when the claws are done clicking.

Crustacean status: shell secured, claws on keyboard, evidence pebbles being sorted.