refactor(frame): dissolve RenderFrameSettings — passes read from state + ctx

docs/BACKLOG.md

@@ -83,6 +83,12 @@ Diagnosed but unplanned. Captured here so they don't get lost; promote to a spec
 - **Milliquas needs its own colormap (AGN ≠ galaxy)** — Milliquas points render overwhelmingly blue. The *clamp* half shipped (#282): the redshift K-correction (`kPerZ`) was subtracting more than the whole `[0,2]` ramp span for high-z quasars and pinning every row to the blue floor; `kPerZ` is now 0 so the real B−R spread survives. What remains is the **semantic mismatch**: quasars genuinely have small B−R, so on the galaxy star-forming↔elliptical ramp they legitimately land in the blue third, but "blue" there means "star-forming galaxy" — the wrong reading for a non-thermal AGN continuum. The fix is to give AGN their own visual encoding instead of reusing the galaxy ramp. Directions to weigh in the brainstorm: (a) a distinct AGN ramp (violet/amber) keyed on B−R so quasars read as a different object class; (b) encode **redshift** instead of colour (z is the meaningful axis for objects spanning the observable universe); (c) tint by the Milliquas **class byte** (Q/A/B/K/N/S) or parent-survey byte — both already on the `.bin`. Likely needs a `colourMode` discriminant on `SOURCE_REGISTRY` + a shader ramp branch; brainstorm → spec → plan, not a drop-in.
 - **Tour feature (full)** — finish the camera tour beyond the Part-2 seed. Tracked design: [splash-screen Part 2 plan](superpowers/plans/2026-05-20-splash-screen-02-stub-tour.md) (ready to execute; engine-side `engine.tour` seed, verified current 2026-06-04) and [2026-05-07 tour-animation spec](superpowers/specs/2026-05-07-tour-animation-design.md) (re-grounded 2026-06-04; labels + MW impostor + engine-API decisions resolved, cinematic palette documented). Execute Part 2 first, then resolve the spec's remaining open decisions (rotation slerp, caption producer, beat list/timing), promote to a plan, and extend the seed into the real waypoint tour.
 - **Label declutter flickers under continuous camera motion (make it a toggle)** — surfaced 2026-06-18 while spiking the cosmic-flows orbit capture. The label director's greedy screen-space overlap cull (`labelDirectorSubsystem.ts` `declutter`, `DECLUTTER_MARGIN_PX = 48`) suppresses the lower-`prominencePx` of any two anchors landing within 48 px in both x and y. Under a slow orbit (or any sustained camera move) labels repeatedly cross that margin and get suppressed-then-released frame-to-frame, which reads as flicker — distracting on a screen recording, and arguably on normal navigation. Two threads: (1) **add a user setting** to disable declutter — belongs under **Settings → Labels → Advanced** (the Labels section already exists; this is a new per-section Advanced toggle, plumbed `settings.labels.declutter` → `labelDirectorSubsystem`). A throwaway `?nodeclutter` URL gate was added in the `fly-to-edge-spike` worktree as a stopgap and should be replaced by the real setting. (2) **Stabilise the cull itself** so on/off decisions hysteresis-damp rather than toggle per-frame (e.g. a release margin wider than the suppress margin, or a short cooldown before a suppressed label can re-show) — the proper fix if labels are wanted *on* during the cosmic-web clip (Post 1) and the tour.
+- **GPU-handle nullability + the `ctx`/`PassDeps` re-threading it forces (revisit)** — surfaced 2026-06-21 while specing the [RenderFrameSettings dissolution](superpowers/specs/2026-06-21-dissolve-render-frame-settings-design.md). **Tackle *after* that dissolution lands** — RFS is the smaller, cleaner first cut of the identical "bag re-threads what's already on state" pattern, and finishing it makes this one obvious.
+  - **The root fact.** Every field on `EngineGpuHandles` is `Renderer | null` because `createEngine` returns synchronously but the GPU pipelines are built in an async `requestAdapter`→`requestDevice`→shader/atlas chain — absent for ~2 bootstrap frames (and re-nulled by `destroy()`). One *transient* lifecycle fact is encoded as *perpetual* per-field nullability, so every access forever pays a null-check tax though the handle is provably non-null after `initGpu`.
+  - **The inconsistency.** Four different access styles coexist for the same kind of thing: (1) **ctx-narrowed** — `renderer`/`postProcess`/`volumeOffscreen`/`texturedDisks` narrowed once via `isEngineReady`/`ReadyFrameContext` (`frameContext.ts:57-72`), read as `ctx.renderer`; (2) **`!`-asserted** at point of use (`pointSpritesPass.ts:98` `state.gpu.focusUniform!`); (3) **null-checked** at point of use (~15 renderers); (4) **always-non-null null-object** (`timingService`'s no-op stub — the better pattern, already in the tree). The `runFrame`→`renderFrame` assembly even pulls renderers from two sources inconsistently (`runFrame.ts:344-348`: some from `deps`, some from `state.gpu`).
+  - **`PassDeps` is the renderer-equivalent of `RenderFrameSettings`.** It re-threads renderers already on `state.gpu` purely to launder the `| null` into a non-null shape at the `renderFrame` boundary — the docblock admits `texturedDiskRenderer`/`proceduralDiskRenderer`/`milkyWayRenderer` live on the bag *only* for `destroy()` reachability and "are not consumed via this bag at runtime (the frame loop receives them through `RunFrameDeps`)." Fix the nullability once and passes read `state.gpu.X` directly (they already get `state`); `PassDeps` sheds its renderer fields. Also folds in the `ctx` half: `ctx` re-declaring `state.gpu.*` handles is the same narrowing-laundering.
+  - **The careful caveat (scar tissue).** The point-of-use checks are *partly deliberate*: the docblocks exclude handles from `isEngineReady` because "bootstrap progression isn't the inverse of teardown" — the 2026-05-08 black-screen incident (memory `feedback_lifecycle_vs_teardown_invariants`), where consolidating a multi-handle "ready" predicate over-constrained mid-bootstrap callbacks and blanked the canvas. So the naive "one big ready flag" is exactly the move that bit us.
+  - **Target.** Separate the three concerns currently fused under one `| null`: **existence** (non-null after `initGpu` — nearly all), **data readiness** (`filamentRenderer` exists but empty until `loadFilaments`; `flowFieldRenderer` until demand — genuinely still absent at draw time), and **destroy reachability**. Narrow the bootstrap-guaranteed handles once into a non-null "ready GPU" view (extend `ReadyFrameContext` or a `ReadyGpu` bag), keep `| null` only for the genuinely-absent-at-draw-time ones (some via a `timingService`-style null-object), let `destroy()` iterate the raw bag — done incrementally, respecting the teardown asymmetry, never a big-bang ready flag.
 - **Thumbnail quality (SDSS / DSS branches)** — the auto-fetched SDSS-cutout and CDS-DSS thumbnails still have the original quality issues: ranked fix options are mask, sky-sub, per-galaxy size, DESI source, brightness norm (see memory `project_thumbnail_quality`). The *famous-galaxy* branch is now fully addressed — procedural-disk fade-out, high-res LOD (#214), and thumbnail calibration + square deproject + disk-plane unification (#229/#234/#235/#240) all shipped — so this item is scoped to the non-curated SDSS/DSS path only.
 - **Supercluster/wall shape accuracy (focus mode)** — cluster-focus mode (PR #242) renders membership as a sphere of radius `apparentRadiusMpc ?? physicalRadiusMpc` centred on the catalog centroid. For superclusters/walls (MSCC) this is crude: the structure is a flattened sheet, so the sphere swallows foreground/background voids and clips the wall's arms (e.g. Hydra Wall reads ~847 galaxies at medium tier). No all-sky per-galaxy membership catalog exists to replace it — redMaPPer/WHL give cluster member galaxies but only in the SDSS footprint; Liivamägi+2012 gives galaxy→supercluster IDs but is also SDSS-limited and threshold-dependent. Investigate a better proxy: (a) **ellipsoid fit** from MSCC member-cluster positions (`memCl` column — data we already have); (b) **density-field membership** reusing the rhizome/MCPM cosmic-web field or DisPerSE filaments (all-sky, same method the literature uses). Option (a) is cheap and immediate; (b) is more principled and reuses existing plumbing.
 - **GLADE shell artifact at ~400 Mpc** — hard depth boundary created by Task 7 abs-mag filter; 3 fix options deferred 2026-05-04. See memory `project_glade_shell_artifact`.