Skip to content

media-v2-results.md

Latest commit

 

History

History
137 lines (111 loc) · 7.16 KB

media-v2-results.md

File metadata and controls

137 lines (111 loc) · 7.16 KB
Raw

Media v2 results

Date: 2026-07-25

2026-07-26 attended-Mac final addendum

The remediated Weaver implementation is now ATTENDED PASS WITH ONE RATE GAP, superseding the round-3 “not yet completed” and “needs attended Mac” statements below. On an M2 Mac running macOS 26.5.1, real metadata, artwork, transport, honest seek/no-session false, helper and player lifecycle, idle-zero, Noro controls/pixels, channel recovery, drag, and persistence all passed. The only attended product check not run was a real payload reporting playbackRate: 2; that behavior remains unit-covered.

The attended run found four live-only defects. The accepted stack now uses short posix.read calls for persistent adapter/command/ack streams, one Zig awake-clock domain for provider watchdog supervision, and a transparent 312x12 Noro seek target over the unchanged 312x3 pixels. The Windows paused pre/fixed A/B had identical SHA-256 and zero differing pixels, and a real pointer click visibly sought the Spotify session. Full evidence is in docs/media-evidence/pr05-mac-verify.md and docs/media-evidence/pr05-visual.md.

2026-07-26 round-3 addendum

The Objective-C helper tests and hosted macOS session passed on the completed exact heads recorded in docs/media-run-status.md; every final implementation head passed the repository gate, both headless Mac jobs, and the hosted Apple-silicon session. Round 3 removes a post-first-frame 100 ms adapter poll and makes a runtime-detected macOS command-send failure process-fatal so host supervision replaces the widget and its authenticated endpoint. The attended fold additionally keeps short command/ack reads prompt while blocking in poll(2) between frames, so ordinary nonblocking EAGAIN is not mistaken for channel loss. The attended real-player matrix is PASS except for the explicitly not-run real 2x playback-rate session, which remains unit-covered.

2026-07-25 round-2 addendum

The second adversarial pass is implemented locally. The accepted state now also includes a deadline-bounded macOS runtime send, race-free Windows reader shutdown, binding- and signature-complete CLI capability detection, crash-supervised recovery from a fatal shared provider channel, a 10-second macOS first-frame watchdog, repeated seek read-back through the two-second deadline, and validated playback-rate timeline math. The Noro seek handler now uses event.u; a live Spotify recheck visibly advanced from 03:24 to 04:09 after a 75% strip press without changing the rendered shell.

The Objective-C helper build/tests and hosted macOS session subsequently passed for the exact round-2 heads recorded in the run status. Those synthetic checks do not verify real-player behavior for the watchdog, non-1× advancement and seek, or fatal-channel restart; those remain UNVERIFIED (needs attended Mac), and the earlier spike still proves only the route.

2026-07-25 adversarial-remediation addendum

The historical macOS BLOCKED conclusion at the end of this file is superseded. The standalone attended spike gate is PASSED per docs/media-evidence/pr04-mac-spike.md. Layer 04 now implements the route and passes the locally executable Windows, portable, release-audit, and macOS Zig semantic gates. The implementation itself remains UNVERIFIED (needs attended Mac); the spike did not run Weaver's provider. Exact macOS 15.4 floor behavior is separately UNVERIFIED (needs attended Mac at 15.4).

The accepted implementation state includes PID-bound endpoints, strict EOF framing, idle-zero transport, exact command deadlines, structurally non-lossy ack/nack lanes, truthful helper failure rejection, read-back-verified seek, timestamped timeline advancement, normalized artwork, retained Windows art on refresh failure, bounded platform waits/teardown, stable restart backoff, and the runtime OS floor. The finding-by-finding evidence and remaining attended matrix are in docs/media-run-status.md.

Windows acceptance result

The Windows slice is complete through the noro gate:

  • examples/noro-shell observes real media and time provider frames.
  • The 188px screen area uses host-cached artwork when available and retains the bundled cover as a conditional fallback.
  • Elapsed time, title, source status, and play/pause glyph are live.
  • Existing previous, play/pause, and next buttons deliver real SMTC commands.
  • The former static 312x3 progress stack is a pixel-matched click-to-seek button using the press event's normalized local event.u.
  • skills/conjure-widget/SKILL.md teaches MediaData.artPath, useMediaTransport, the media-transport capability, and promise semantics.

The full viewed visual checklist and capture inventory are in docs/media-evidence/pr05-visual.md.

Live Windows result

Spotify was used as the real player. Art and metadata were already visible on the first settled provider frame. Pause/play visibly changed the glyph and timeline, next changed art/title, and previous returned to the prior track.

Seek was measured with a temporary diagnostic title. The track duration was 04:16 (256 seconds); a click at 75% targeted 192 seconds and the next provider frame displayed 03:12 (192 seconds). The diagnostic was reverted before the final capture and commit.

The installed Rainmeter original was captured beside Weaver. Rainmeter was in its genuine standby state, while Weaver was connected to Spotify; the evidence therefore supports shell geometry/styling parity and Weaver's live state, not a claim that both processes rendered the same media frame.

Idle A/B

This is a controlled source-only A/B on the same machine and same media-v2 ReleaseFast runtime/host binaries. The baseline used examples/noro-shell/widget.tsx byte-for-byte from master; the candidate used the PR 05 source. Each process set settled for 40 seconds, then process CPU, private bytes, and thread counts were sampled across approximately 60 seconds. Spotify was paused and stable.

Metric Master noro source Media-v2 noro source Difference
Sample 60.027s 60.039s +0.012s
Widget CPU 0.000ms 562.500ms +562.500ms
Host CPU 218.750ms 359.375ms +140.625ms
Combined CPU 218.750ms 921.875ms +703.125ms
Combined CPU / one logical core 0.364% 1.535% +1.171 percentage points
End private memory, widget + host 41.81 MiB 47.12 MiB +5.31 MiB
End threads, widget + host 8 17 +9

Performance claim: explicitly declined. This single matched run measures the cost of activating the Windows media provider, dynamic artwork path, timeline updates, and transport endpoint. It does not support a no-regression or broad benchmark claim. It does show bounded steady-state behavior over the measured minute: candidate private memory changed by about 0.03 MiB during the sample.

Historical macOS acceptance (superseded by the addendum above)

The original Windows-only run recorded PR 04 as spike-gated and BLOCKED because it could not produce a real macOS metadata frame plus delivered command. That historical decision is retained for provenance only. The later attended route evidence passes the spike, while Weaver's remediated implementation and the macOS noro side-by-side remain unverified as stated in the dated addendum.