[alfredodeza]

Summary

Closes #3601.

apr inspect (via RosettaStone::inspect → GgufReader::get_tensor_raw)
refused any tensor carrying an IQ-family or ternary ggml dtype with
Unsupported dtype N for raw extraction, even though aprender-serve's
GGML_TYPES table (#3432) already has correct, vetted block/byte sizes for
these ids. Real Unsloth UD-* dynamic quants mix IQ4_XS/IQ2_XXS/etc.
into otherwise-K-quant GGUFs, so real downloads crashed apr inspect
(repro: Qwen3.5-4B-UD-Q4_K_XL.gguf, dtype 23 = IQ4_XS).

Change

• crates/aprender-core/src/format/gguf/shape.rs: add the missing ggml ids
  (16-23 IQ2_XXS..IQ4_XS, 29 IQ1_M, 34/35 TQ1_0/TQ2_0) to get_tensor_raw's
  byte_size match, using the same blck_size/type_size pairs as
  aprender-serve/src/gguf/ggml_type_table.rs's GGML_TYPES so the two
  tables agree.
• crates/aprender-core/src/format/gguf/builder.rs: regression test
  test_get_tensor_raw_sizes_every_iq_and_tq_id builds a synthetic
  single-tensor GGUF per new id and asserts get_tensor_raw returns the
  exact byte count instead of refusing.

This is a pre-#3421/Phase-2 leaf fix (same shape as #3432): aprender-core
can't depend on aprender-serve, and the unified TRAITS crate #3430
proposes doesn't exist yet, so this duplicates the vetted sizes locally.
When Phase 2 lands, this call site becomes one more to migrate.

Test plan

• cargo test -p aprender-core --lib format::gguf — 380 passed
• cargo clippy -p aprender-core --lib -- -D warnings — clean
• Manual repro on the pinned release binary: apr inspect ~/Downloads/Qwen3.5-4B-UD-Q4_K_XL.gguf now succeeds (previously
  error: ... Unsupported dtype 23 for raw extraction)

🤖 Generated with Claude Code

[alfredodeza]

Superseded by #3660, which lands the same fix more completely:

• PMAT-3601/PMAT-3656: apr inspect/tensors read IQ/TQ GGUFs — aprender-core sizes ggml types from TRAITS, and refuses IQ dequant instead of inventing weights #3660 routes through trueno_quant::TRAITS (PP-QUANT-001 M1, merged in
  PMAT-3430 M1: one ggml tensor-type table, extracted from upstream rather than typed #3581 — landed on main after this branch was cut, hence the collision).
  This PR added the missing IQ/TQ sizes as a fifth hand-typed copy in
  shape.rs, which is exactly the duplication class ARCH: quant-type dispatch is duplicated across ~30 files with no single source of truth — propose a dedicated consolidation release #3418/EPIC: PP-QUANT-001 — quant-type dispatch, Phase 0/1 (0.69 train) #3421 exist to
  eliminate. PMAT-3601/PMAT-3656: apr inspect/tensors read IQ/TQ GGUFs — aprender-core sizes ggml types from TRAITS, and refuses IQ dequant instead of inventing weights #3660 is objectively the better fix now that the unified table
  actually exists.
• PMAT-3601/PMAT-3656: apr inspect/tensors read IQ/TQ GGUFs — aprender-core sizes ggml types from TRAITS, and refuses IQ dequant instead of inventing weights #3660 also fixes two bugs this PR didn't touch: safetensors.rs's
  ggml_dtype_element_size had 7 wrong IQ sizes and a misordered tail
  (falsely refusing a valid UD-IQ2_XXS file as "Truncated"), and
  get_tensor_f32's dequantize_iq_approximate was silently fabricating
  garbage weights instead of refusing (apr convert/import invent IQ-quant weights: get_tensor_f32 'approximates' IQ2..IQ4 from raw bytes and exits 0 (std 36.6× the real tensor) #3656) — worse than the crash this
  issue reported.

Closing this one; #3660 is the fix that should land for #3601.

[github-actions]

§13.11 rung 1 — quorum shadow verdict

S13-SHADOW pr=3654 head=f60304b3c178d716c03c12972c8f789e7b0ba938 verdict=REFUSE class=Q1 arm_rc=1

Shadow mode: this records a verdict and merges nothing. A refusal
to arm is not a block (§13 adds zero rows to §7) — the pull request is
exactly as green as it was.