fix(adr-025): inclusive acceptance (≤ τ) makes the LOD formula exact at the boundary

[AdaWorldAPI]

Summary

Closes Codex P2 on PR #45 (now merged). Fresh PR rebased on current main (includes #45).

The edge case

ADR-025 §2 derived the level formula from a strict acceptance predicted_error_L < τ but stated the result as r* = ⌈2·log_b(C/τ)⌉. These disagree at the exact-power boundary C/τ = b^k:

• strict < needs ⌊2·log_b(C/τ)⌋ + 1 = 2k + 1
• the doc wrote ⌈2·log_b(C/τ)⌉ = 2k

At r* = 2k the predicted error is C·b^{-k} = τ exactly — equal to tolerance, not strictly below. So the ⌈⌉ formula selects one level too shallow at that boundary.

The fix

Change the acceptance to inclusive (≤ τ — error within tolerance), which is also Cesium SSE's own convention (sse ≤ maximumScreenSpaceError, already referenced by crates/cesium/src/sse.rs):

C·b^{-r/2} ≤ τ  ⟹  b^{r/2} ≥ C/τ  ⟹  r ≥ 2·log_b(C/τ)
⟹ r* = ⌈2·log_b(C/τ)⌉                                (now exact)

At C/τ = b^k: r* = 2k, predicted_error = τ, accepted by ≤. The ⌈⌉ formula is the smallest integer r ≥ 2·log_b(C/τ), exactly right under inclusive acceptance.

For deployments needing strictly below tolerance, the doc now gives the alternative explicitly: r* = ⌊2·log_b(C/τ)⌋ + 1 (one level deeper at the boundary).

The decision diagram's predicted_error_L < tolerance → ≤ tolerance for consistency.

Review (numerically verified)

C/τ = b^k (b=16)	inclusive r*=2k → err/τ	strict r*=2k+1 → err/τ
k=2 (256)	4 → 1.0000 (accepted by ≤)	5 → 0.2500
k=3 (4096)	6 → 1.0000	7 → 0.2500
k=5 (1.05e6)	10 → 1.0000	11 → 0.2500

The canonical anchor r* = ⌈log₄(C/τ)⌉ (the b=16 reduction) and the "5-12 hop" range are unchanged — this only tightens the acceptance predicate so the existing ⌈⌉ is exact rather than off-by-one at powers of b.

PII abort-guard (word-boundary): CLEAN. cargo check: clean (docs-only).

https://claude.ai/code/session_01PBTGaPCSnnt6u3pjXpbLwY