[m5agent]

Summary

The browser could run reference firmware, but lacked a validated way to edit and run UIFlow code across its admitted devices. This adds a default UIFlow2 programming environment and MicroPython editor for all 19 browser-runnable devices.

• Compose each official base with a deterministic startup and verify file readback.
• Execute pasted/imported Python through the guest UART0 raw-paste/raw-REPL path, with actual output, pause/resume, interrupt and hardware-reset recovery. UART ownership/console epochs fence stale executions; repeated interrupts send one running Ctrl-C.
• Validate model buttons, keyboard and body touch through guest controllers/drivers and bind one-instruction execution to 11 exact S3 workbench Flash/main variants. The pinned WASM rebuild includes bounded host-poll wakeups.

Validation

• Virtual-device validation: canonical V6 R8 passed 60/60 cases, 2,564 checks: 19 editors, 19 button cases, Cardputer keyboard, StackChan body touch and 20 fixed examples. This includes 224 raw-paste transfers, 38 untouched default-source runs and 200 input trials.
• Final source: c3db057d2ee99c022a4116ae72d661c4d39888c6. Exact firmware/runtime/policy and cleanup evidence. The final head preserves all 49 tested source-vector files byte-for-byte.
• Final host R2: 1,957 passed / 10 optional skips / 0 failures out of 1,967 Python tests; 28 Node contracts, 9 collector controls, four validators and diff/scope checks passed. R1's three documentation failures and targeted 21-test repair remain recorded.
• Ordinary source CI: both jobs succeeded; this is separate from the manual browser gate.
• Published work record.

Scope and remaining limits

Policies require their exact Flash, startup source, fixture, ready marker and patch. Unknown firmware and fixed/classic images retain separate contracts. RUNTIME-07/08/09 remain OPEN; policy passes do not prove a shared C root-cause fix or long-term stability.

PM1 acceptance uses explicit application configuration: a 500 ms click window with unchanged 250 ms input. StopWatch/StickS3 use their public driver; PaperColor/PaperMono explicitly use public I2C RMW. Default-mode PM1 events are not established. Cardputer Fn pointer input includes a grave-key companion. Missing held-iframe CDP fences remain NOT CAPTURED.

Independent local cold samples on the same original fixed images changed from 109.458 to 54.775 s (Matrix) and 90.687 to 39.144 s (Plus). These are source-pinned, single observations for those two images.

Four original-image cold-prime → retained-page warm pairs completed: 53.491 → 76.480 s (75 Matrix), 39.378 → 63.627 s (75 Plus), 109.448 → 108.710 s (3c Matrix), 96.703 → 111.456 s (3c Plus). Warm includes model reconstruction; pure-cache and general speed guarantees remain unestablished. The first observer failure is retained. The two-guest functional gate is not a performance benchmark.

SDK pthread cancellation remains unestablished. Main merge, Pages publication and online acceptance are NOT RUN and await approval; physical equivalence is not claimed.