feat(bgz-tensor): HhtlF32Tensor codec + Path A encoder + Path B argmax-parity probe

[AdaWorldAPI]

Summary

Two follow-up paths from PR #183's reconstruction failure, both tested on real Qwen3-TTS-0.6B.

Path	What	Status
A	HhtlF32Tensor — f32 CLAM centroids replace Base17 palette; keeps SlotL on top	Running — partial results ρ̄ ≈ 0.2–0.5 on attention/embed tensors
B	cascade_attention_probe — argmax-parity probe on single attention head using existing Base17 palette + FisherZTable	FAIL — 3.71% top-1 agreement (19/512)

Path B finding (architectural)

Path B reused the existing Base17 palette as the indexing substrate. Argmax parity vs raw f32 Q · K^T collapsed to 3.71%. Root cause is the same class of error as #183: Base17 fold discards direction information, so two rows that are near-neighbours by f32 dot product land in completely different palette cells.

Paths A and B are NOT separate tracks. Path B's FisherZTable approach only works if the palette faithfully partitions the row space by inner-product similarity. Base17 palette doesn't. f32 CLAM palette might. Path B depends on Path A's f32 palette substrate before it can work. This is a session insight that wasn't clear before running the probe.

Path A preliminary numbers (run still live)

[1/26] ARGMAX code_predictor/down      [1024×3072] × 5   ρ̄=0.27
[2/26] INDEX  code_predictor/embed     [2048×1024] × 15  ρ̄=0.47
[3/26] ARGMAX code_predictor/gate      [3072×1024] × 5   ρ̄=0.21
[4/26] INDEX  code_predictor/lm_head   [2048×1024] × 15  ρ̄=0.41
[5/26] ARGMAX code_predictor/qko       [1024×1024] × 5   ρ̄=0.49

~10× better than HhtlDTensor's ρ ≈ 0.04 on the same tensors — confirms the Base17 reconstruction pathology and that f32 centroids are moving the right direction — but still short of the 0.95 argmax / 0.98 index targets.

Why Path A isn't hitting targets

On 1024-dim near-orthogonal rows:

• k=256 centroids: nearest centroid to any row is at L2 distance ≈ √2 × row_norm (near-orthogonal to target)
• 8 SVD coefficients recover < 1% of signal direction for rank-1024 rows

Two clean fallbacks the other session's review named (E4, E5):

• k=1024 or 2048 centroids — larger palette, denser coverage
• Per-leaf local SVD basis — 16–32 components per centroid cluster instead of 8 shared

Full run numbers + decision will follow in a comment when Path A completes.

What this PR contains

File	Purpose
crates/bgz-tensor/src/hhtl_f32.rs (+316 L, 5 tests)	New reconstruction-grade codec: f32 palette + SlotL
crates/bgz-tensor/src/lib.rs (+1 L)	pub mod hhtl_f32;
crates/thinking-engine/examples/universal_hhtl_f32_encode.rs (+264 L)	Path A encoder with regime dispatch, gates 1 + 3
crates/thinking-engine/examples/cascade_attention_probe.rs (+192 L)	Path B argmax-parity probe (single attention head)

Tests passing

test hhtl_f32::tests::entry_byte_size_is_one ........................... ok
test hhtl_f32::tests::encode_without_leaf_picks_real_rows_as_centroids . ok
test hhtl_f32::tests::reconstruct_without_leaf_returns_nearest_centroid  ok
test hhtl_f32::tests::storage_accounting_is_additive ................... ok
test hhtl_f32::tests::encode_with_leaf_beats_without_leaf_on_real_rows . ok
5 passed; 0 failed

Recommendation

Merge the module + tests even if Path A gates fail on full run — HhtlF32Tensor is a correct primitive. The example can be either:

1. Merged as-is with a note documenting the measured ρ̄ ≈ 0.3 result, OR
2. Closed, reworked in a follow-up PR with k=1024 + per-leaf local SVD

Depends on full run result — will update this description when it lands.

Session PR stack

PR	Status
#180	merged — SlotL foundation
#181	merged — HhtlDTensor × SlotL
#182	merged — SharedPaletteGroup × SlotL
#183	merged — universal_hhtld_encode (Base17 path, reconstruction failure documented)
#184	this PR — HhtlF32Tensor + Path A/B probes

https://claude.ai/code/session_01NYGrxVopyszZYgLBxe4hgj

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 16ad36001c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Reject palette sizes that exceed 255 centroids

HhtlF32Entry.twig is a u8, but encode accepts any k and assign_nearest_f32 stores centroid IDs via ci as u8. When k > 256, centroid indices above 255 wrap (e.g., 300→44), so rows are silently assigned and reconstructed from the wrong centroid, corrupting cosine/argmax quality measurements instead of failing fast. Please clamp or reject k beyond the representable range (and validate k > 0) before writing twig indices.

Useful? React with 👍 / 👎.