[brsbl]

Generated choice, string and boolean controls could accept input without changing their output. The parameter bridge now accepts primitive choice arrays and string/boolean declarations, preserves value types and numeric bounds, and ignores its own queued state messages in standalone previews. Regression tests exercise the actual bridge.

Quick explanations load a short workflow and parameter example, publish before one browser check, and keep the full Figure/Hero workflow behind explicit deeper requests. The launch observer now proves painted iframe content, checks frame identity/layout across capture, and wakes lazy frames before waiting for load. Screenshot-completion time is a conservative visible-card upper bound; status readouts count as output, while cosmetic controls still fail.

Final-head launch eval

All 20 prompts per provider ran once in fresh non-Git Personal threads on default models at 2ff907cc5adc65a16d6efc9974b2e2d08e5cd9a5. All 40 delivery screenshots and 63 automatically failed control verdicts were reviewed; real-input checks resolved harness gaps without rerunning agents or changing measured times.

Provider	Median / p90	Every control live	Clarifying questions	Build progress
Claude Code	2.47 / 4.57 min	18/20	0	19/20
Codex	1.41 / 1.68 min	19/20	0	20/20

Both meet the stated launch bar (≥18 live, no questions, median ≤3 min, p90 ≤6 min). The existing strict harness still fails Claude's all-20 progress requirement: one settings answer showed only a final preface. Remaining generated-output failures are Claude's washer Replay and database Run query under reduced motion, and Codex's cosmetic settings navigation. These are retained failures, not waived controls.

Gate-1 reference: Claude 3.93/5.52 min, 19/20 live; Codex 1.46/1.70 min, 15/20 live. Claude first-tool→first-browser mean is 98 s across 18 mechanism figures (previously 171 s across 17). Current mean elapsed envelopes: startup 7 s, pre-publication 91 s, publication→browser 7 s, browser→visible 74 s. Different subsets, shared-host conditions and stronger capture proof prevent a controlled causal claim. Eight runs used the fixed 30-second render grace for 11–509 ms; all are logged within the unchanged 15-minute cap.

Validation

• Remote CI: 41 successful checks on the final head, including Playgrounds and install-release validation (run).
• Full cumulative Slop Cop: no unresolved P0/P1. The lazy-frame deadlock found in review was fixed and revalidated on this head. P2 follow-ups: sparse overlay sampling (all 40 captures inspected), a conservative grace-deadline race, inherited TMPDIR (explicit worktree alias supplied), predeclared numeric select serialization, and direct full-mode setup guidance.
• Isolated bb E2E, Chrome for Testing: bridge input/reset/state persistence, 20 control fixtures, 7 visibility fixtures, full 20×2 eval, and deeper-reference routing passed. Source/server identity was checked before QA writes; every QA CLI names the isolated URL. No production configuration or data was used for QA.

Bridge regression screenshots

Before: 7be5770a6c552eb4b5632ec11b259ef52f3700d5 · After: 2ff907cc5adc65a16d6efc9974b2e2d08e5cd9a5 · Original 2× PNGs, identical fixture and input sequence. Before stays at choice a / message Ready; after reaches c / Changed with a boolean true and changed output geometry.

State	CSS viewport	Before	After
Changed controls	390×900
Changed controls	1440×900

Closes #519

BB-Thread-ID: thr_vv8zgxmczt

AGENT GENERATED

Post-merge: main CI passed. Published release f6818ad879ec2d29cf3df1d138b1c285c6389199 has 85 shipped files that match the release artifact.