chore: remove the quality-tier API nothing drives

[N1k4G]

Closes #128. Adaptive quality is not planned, so per that issue's decision the dead API goes rather than being wired.

What was dead

setQualityTier() and placementCount had no callers anywhere in src, tests or scripts, and neither was on the SceneRenderer interface, so nothing could reach them polymorphically either. #qualityTier was initialised to "high" and never changed, so the tier path in the renderer never executed.

Left in place that is worse than no API. A reader finding setQualityTier() reasonably concludes quality adapts — it does not. placementCount documented two consumers, "tests and the diagnostics overlay", neither of which exists.

What stays, and why

The tier machinery underneath is real and works:

Tier	Assets	Dropped below it
low	18	— cars, lorries, containers, engine, helm, cave columns, anchor, warning sign
medium	+10	towels, umbrellas, soft corals, gorgonians, light shafts, bunks, netting
high	+1	the reef fish shoal

Every asset declares a minimumQualityTier, buildSceneLayers filters on it, and tierAllows is tested as a pure function. Callers just always pass "high". That is now a named constant with the reasoning attached, so whoever wires adaptation later finds the filtering already built and tested, and only needs something that decides a tier — a frame budget, a device probe, or a user setting.

The tiering is also sensibly ordered, which is worth preserving: things a diver navigates by survive, decoration goes first. Dropping the wreck to low removes dressing but leaves every structural landmark.

Record corrected

docs/decisions.md listed quality tiers among the PixiJS scene's delivered properties. Removed, so the record matches the code.

The mention in docs/baseline/README.md stays — that is a field to record when benchmarking a device, not a claim that adaptation works.

Verification

typecheck, lint, sites:check, 133 unit, 28 parity, 36 e2e.

The Pixi visual guard reports 0% changed pixels on all three scenes, which is the expected result of changing how an unchanged value reaches the layer factory — and the first time that guard has been used to confirm a refactor is visually inert rather than to catch a regression.

[N1k4G]

Reviewed 0b58be3: no findings. The dead setter/getter and mutable field are removed without disturbing marker pooling; the layer factory remains explicitly pinned to the tested high tier; and the delivered-properties record now matches the product decision. Pixi visual guard: 0% changed pixels on all three scenes. Focused asset/layer tests: 38/38 passed. Current CI is green.