feat(hud): put the verdict's explanation first, make Slim a MangoHud bar, add the network row

[papi-ux]

Summary

• Command Center opened on a one-line verdict and buried what explained it. The session strip says how the stream is doing; the Doctor's reading and the Stream card that say why, and what Auto is running, sat below Quick Keys, Controls, Session, and Overlays, three screens down on a Retroid. They now come right after the strip. Overlays follows because the HUD switch is the frequent tap, then Controls and Session, then the full Quick Keys grid, then Sync and Advanced.
• Esc, Meta, and Alt + Enter stay pinned in one row directly under the strip, so reaching the game's own menu on a handheld never needs a scroll. The pinned row reuses the grid's own action instances: the grid stays the whole keyboard and both rows fire the same key.
• This reverses the 1.4.3 order deliberately. Keys-first assumed the strip was enough, and it is not when the number is bad.
• NovaHUD Slim is now a one-line bar in the MangoHud spirit: the health bar, the frame rate with its last minute drawn beside it, then decode time, round trip, and bitrate with 7 sp inline labels, so two millisecond values are never told apart by position. Bitrate drops its unit like Performance. Minimal stays the sparse two-line card, so Slim is about height, not about hiding data.
• NovaHUD Debug gains a network row. The decoder has reported packet loss, round-trip variance, and the session's lost-frame count in every sample and no HUD mode showed them. LOSS is graded so zero is the only green: a trace reads "<0.1%" in warning rather than rounding to a green-looking 0%, under one percent is warning, one percent and over is danger. JIT reads "--" while there is no round trip to wobble around. DROPS is the session's cumulative count.
• Source guards re-pin the new first-paint order, the Slim composition, and the Debug network row. Unit tests cover the pinned-key selection and identity, the Slim bitrate form, and the loss grading and formatting edges.

Exact candidate

• Commit: 32ebb14
• Tree: 665527e18cfc6067c49c33c49736fdb3c404a5d5
• Parent: bcaaede
• Base: bcaaede (master)

Verification

• :app:testNonRoot_gameDebugUnitTest (x86_64): 1617 tests, 0 failures, 0 errors. Note: pc-papi's /tmp is a tmpfs with a per-user quota that other sessions' Polaris asan/tsan build trees have filled; Robolectric and Mockito extract into the Java temp dir, so the suite fails there with "Disk quota exceeded" regardless of the code. The green run used an init script that points the Test tasks' java.io.tmpdir under /home. Nothing in the repo changed for that.
• :app:lintNonRoot_gameDebug -PlintFailOnError=true: clean
• :app:assembleNonRoot_gameDebug (arm64-v8a), installed on the Retroid Pocket 6 as com.papi.nova.debug
• Device, Command Center: first paint is header, session strip, the pinned Esc / Meta / Alt + Enter row, the Doctor card, the Stream card with its Launch preset; scrolling continues Overlays (HUD switch, mode picker, Stats Overlay, Copy Diagnostics, both opacity rows), Controls, Session, the full Quick Keys grid, Sync Status, Advanced.
• Device, Slim: one pill, health bar, fps, sparkline, DEC 6.2ms, RTT 3ms, BIT 17M. Guide + Y still walks Slim, Minimal, Performance, Debug, Slim, and the mode persists.
• Device, Debug: fourth row LOSS 0% (green), JIT 0ms, DROPS 2 under HOST / IN / OUT.
• dumpsys gfxinfo, 60 s windows on the RP6. The HUD recomposes once per sample, so these are the UI thread's one frame per second, not video frames. Control would not come up on the host this evening (the session showed the host desktop at 12 fps), so this PR's rows are from that desktop session; the reference rows from feat(hud): decode time in Debug, a Slim layout, and a Command Center that stays out of the way #283 were taken inside Control at 60 fps. The UI-thread work is the same either way.

HUD layout	Session	Frames	Janky	p50	p90	p99
Slim, this PR, window 1	desktop, 12 fps	59	2	11 ms	13 ms	14 ms
Slim, this PR, window 2	desktop, 12 fps	59	9	12 ms	13 ms	15 ms
Debug, this PR (4 rows), window 1	desktop, 12 fps	59	11	12 ms	14 ms	16 ms
Debug, this PR (4 rows), window 2	desktop, 12 fps	59	13	12 ms	14 ms	15 ms
Debug, #283 v2 (3 rows), reference	Control, 60 fps	59	11	12 ms	13 ms	16 ms
Performance, #283, reference	Control, 60 fps	59	0	11 ms	12 ms	13 ms

Read: Slim with its sparkline lands where Performance already was, 11 to 12 ms p50. Debug's fourth row is inside run-to-run noise against the three-row version. Janky counts move by several frames between identical windows, so treat those as a range, not a delta.

Not covered

• androidTest is not compiled by CI and already fails to compile on master for unrelated reasons; the HUD androidTests were not touched or run.
• The companion-display Command Center is still the legacy GameMenu, as before.
• MangoHud draws per-frame frametimes; the HUD's samples arrive once a second, so the Slim sparkline is a 60 second trend of the frame rate, not per-frame jitter.
• No 60 fps game capture of the new Slim bar this evening; the desktop session capture shows the same layout with a flat line at 12 fps.