ETK 0.4.0 — Tune the driver, photograph the crash

[0.4.0] - 2026-06-18

Tune the driver, photograph the crash. The hunt for ROCKNIX's headline instability — the a6xx GPU-fault freeze — produced two operator-facing instruments. A new DRIVER tab exposes the Mesa/Turnip dials the crash signature points at and stamps every session in the ledger with the exact dial set it ran under, so genuine-play tuning finally yields attributable data instead of N=1 guesses. And a crash-cam turns every recoverable freeze into a photographed, dial-tagged ledger entry — the frozen frame is grabbed at the R3 panic, bound to its session, and previewed full-screen on the device. Plus the Manage Shaders engine now deploys reliably. The throughline holds: ETK ships tooling, never bytes.

Added

• DRIVER tab — Turnip dials, ledger-tagged (Pitstop). A fifth tab exposing the Mesa/Turnip environment knobs as gamepad dials: TU_AUTOTUNE_ALGO (the GMEM↔system-memory render decision engine) and the TU_DEBUG isolation ladder (nolrz/noubwc/sysmem/gmem + an Advanced group). APPLY injects via the proven profile.d path (097-etk-turnip-dials) — effective next launch, survives a cold boot; Reset reverts to Turnip's built-in autotune. Every APPLY writes active_tune.txt, which session_postmortem records as the ledger's tune_tag column — so each race is attributable to the exact dial set it ran under. One knob per soak, on-screen.
• Crash-cam — frozen-frame capture + on-device preview, bound to the ledger. The dominant ROCKNIX crash leaves no core and no RPCS3 fatal — only a frozen screen. recovery.sh now grabs that frame via grim at the R3 panic (best-effort, hard-timeout-bounded so it can NEVER stall the nuclear recovery), and session_postmortem binds it to the crash's ledger row (crash_shot column). In the Pitstop crash-detail card, press ↑ to view the frame full-screen (via swayimg), dismissable with any button. A crash entry is now signature + frame + dial, all linked.
• Manage Shaders deploy fix. tools/vault_sweep.sh — the engine the Manage Shaders screen drives — is now deployed to the rig by install.sh. It had only ever reached the rig via a dev-time push, so a plain uninstall/reinstall cycle silently dropped it and broke the screen with a misleading "no boundary." The screen now distinguishes engine-missing from a genuinely absent rebuild boundary, and Sweep / Delete-vault / Clear-cache are field-confirmed.
• Busy-frame throbber (Pitstop). TOOLS shader scan/clean and PADDOCK sync/push/pull now animate a ROCKNIX-style ASCII spinner on a background thread instead of freezing on a single frame.

Known issues / deferred

• The a6xx GPU-fault freeze remains the headline instability (carried from 0.3.1) — and 0.4.0's new instruments sharpened it without yet beating it. On-rig DRIVER-tab A/B established the fault is a NULL texture/vertex descriptor (iova=0x0, source=TP|VFD) on the live-race render path (the High Speed Ring tunnel), not memory pressure or tiling — so the TU_AUTOTUNE_ALGO dials (gmem/sysmem) do not move it (4/4 froze). A first probe of RPCS3 Write/Read Color Buffers (the GT6 render-correctness knob) did not eliminate the freeze either, but indicatively moved the fault deeper, past the tunnel-mouth transition (N=1, suggestive). All of which reinforces the Stage-IV Turnip fork as the real next lever — the residual fault sits below the RPCS3 render-setting layer.
• mako cannot render images on ROCKNIX — the gdk-pixbuf loader modules are stripped from the build, so a notification renders text but never an image. The crash-cam preview therefore uses swayimg; a build lacking it degrades to a text toast with the SMB path.