docs(adr): add ADR-024 — Palette256 + HHTL codec as universal compression primitive

[AdaWorldAPI]

Summary

New ADR-024 in docs/ARCHITECTURAL-DECISIONS-2026-06-04.md (+180 LOC). Names an existing primitive — three independent palette256 deployments + one empirical anchor (ρ = 0.9973 vs cosine in lance-graph-arm-discovery) — rather than proposing one. Companion to ADR-022 (The Firewall) and ADR-023 (IR-as-wire-truth).

Per the runtime session's three explicit tightenings:

1. Empirical anchor pinned (ρ = 0.9973 vs cosine — the falsifiable stake)
2. 256-ceiling escape hatches in body (per-tile / hierarchical / palette-64K — pre-empts reviewer question)
3. Cross-references to FINDING-grade anchors (EPIPHANIES.md:28 + old-stack-capability-parity.md §3.39 — ADR names existing primitive)

The codec

HHTL prefix  (NiblePath / quadkey / class identity)
     ↓ establishes spatial / semantic frame
within-frame values cluster
     ↓ quantize to 256-index palette
     ↓ const-table lookup (compile-time HHTL where possible)
1-byte index per element, sub-µs decode, zero heap allocation

Three already-deployed instances (verified by the runtime session)

Domain	Receipt
Security mesh	Binary16K = [u64; 256] @ MedCare-rs/.../graph_contract.rs:36 + Hamming popcount on _effectiveReaders; production wire-up at column_mask_bridge.rs → medcare-server/state.rs:167, 265, 439
Attention	bgz-tensor::WeightPalette::build(…, 256) + AttentionTable::build @ crates/bgz-tensor/examples/compare_stacked_vs_i16.rs:90-92
Distance	lance-graph-arm-discovery aerial codebook — ρ = 0.9973 vs cosine (measured, FINDING-grade)

Runtime side's BindSpace dissolution work (bardioc #18 / lance-graph #470) hinted via the Quintenzirkel qualia codebook (8 B → 1-2 B per row).

Adoption checklist for new domains

1. Identify the prefix (NiblePath / quadkey / class identity)
2. Identify the palette domain — which values cluster?
3. Build the palette + measure ρ-vs-reference. Target ≥ 0.99 to match the arm-discovery anchor.
4. Decode = const-table lookup (zero-allocation hot path).

ρ-vs-reference becomes the adoption contract — reported once per domain at adoption time. Keeps the empirical floor honest as the primitive spreads.

256-ceiling escape hatches (in the ADR body)

• Per-tile / per-frame palettes — cheapest; matches the quadkey-prefix discipline
• Hierarchical palettes — coarser at higher levels, finer per leaf (mirrors the tile pyramid + SH L0/L1 vs L2/L3 split in splat-fit)
• Palette-64K upgrade — 2-byte index; reserve for measured cases (not speculation)

Two next-domain adopters queued

D-id	What	Source
D-OSM-2	OSM tag palette + tile-local coordinate quantization	lance-graph PR #473 (forthcoming cesium-osm-substrate-v1.md); ρ-vs-reference reported on first per-country PBF run per runtime-session's §11 commitment
D-SPLAT-4	SH-aware palette extension on the Gaussian3D carrier	splat-native arc

The falsifiable property

"The same compile-time HHTL prefix + palette256 codec decodes (a) _effectiveReaders for row auth, (b) OSM way attributes at zoom-21 tile, and (c) FMA-bone SH coefficients at sub-microsecond per element with zero heap allocation."

If that property holds across all three, the substrate is doing its job. If it fails on one, the substrate is leaking dialect into the codec.

Cross-arc reuse — sharpens the substrate-reuse claim

The docs/RDF-OWL-ALIGNMENT.md §10 "geographic litmus complements anatomical" framing cashes out as: FMA-bones and OSM-vectors use the same codec (palette256 + HHTL prefix), not just the same IR. The DOMAIN-INSTANCES.md §2.6 callout queued for the post-lance-graph-#473 docs PR will reference ADR-024 as the falsifiable property.

Verification

• cargo check --workspace --all-targets → clean (docs-only PR)
• PII abort-guard (word-boundary): CLEAN

Runtime-side commitment

Once this merges, the runtime session will file a small §11 callout to lance-graph PR #473 (cesium-osm-substrate-v1.md) tying D-OSM-2 + D-SPLAT-4 to this ADR with the ρ-vs-reference measurement as the falsifiability test. That closes the loop from the runtime side — the falsifiable property is testable at D-OSM-2 implementation time, not deferred.

https://claude.ai/code/session_01PBTGaPCSnnt6u3pjXpbLwY