IQ1_S/IQ1_XXXS decode: three parameter defects pass the whole gate (self-referential reference decode)

[localai-bot]

Found by an independent review of PR #946. The author self-reviewed and merged it; these were missed. Row ENG-EXPERT-STREAM, spec .agents/specs/expert-streaming.md.

F1 (HIGH): every decode parameter except the grid is pinned only by self-consistency

ReferenceDotF64 (tests/vt/test_ops_quant_dot.cpp:231-236), which the case comment calls "the INDEPENDENT reference", decodes the weight with vt::cpu::BlockToFloat(wtype) — the very function under test. The MatmulBTQuant NMSE reference does the same (:915). Both are independent only in the SUMMATION, never in the DECODE.

Three mutations, each applied and compiled, are UNCAUGHT by the entire gate with assertion counts unchanged:

Mutation	Meaning
kIq1sDelta 0.125F -> 0.25F	2x error in the reconstruction offset, BOTH encodings
IQ1_S delta sign inverted (dequant + vec_dot)	qh bit 15 means the opposite of upstream
IQ1_S scale from qh bits 13-15 instead of 12-14	off-by-one field position

The grid table IS sealed (FNV-1a digest + lane census) and that seal WORKS: corrupting a grid entry is caught. The seal simply stops one table short of the parameters around it.

Failure mode: silently wrong tokens from a 2.4 T model with every gate green.

F2 (MEDIUM-HIGH): the 2e-3 NMSE ceiling hides a real decode defect

tests/vt/test_ops_quant_dot.cpp:148,153. Measured: unmutated max 5.24e-4 (iq1_s, m=4, n=1); with delta=0.25 it is 6.97e-4.

Ceiling	Unmutated	+ delta 0.25
6e-4	GREEN	RED
2e-3 (as merged)	GREEN	GREEN

The measurement required 1.05x; the ceiling was set at 3.8x. Violates AGENTS.md: "Never make a red gate green by deleting an assertion or widening its scope." The figures quoted in the merged commit (3.8e-4 / 3.8e-5) do not reproduce against the tree.

F3 (MEDIUM): the loader expand path refuses 19 and 66 although vt decodes them

src/vllm/model_executor/model_loader/gguf_dequant.cpp:107-116 / :181 has no case 19 / case 66. RouteGgufTensor routes to kExpandBf16 whenever cpu_ref is set (the oracle switch), keep-quant is off, K is ragged, or the role is not verbatim, so the reference lane cannot load the target checkpoint. ggml 18 (IQ3_XXS) has the same pre-existing omission. Violates "Nothing lands dead" and the refuse-by-name rule.

Lower

• F4: W9's declared gate slice says "CPU tests green against upstream-derived vectors"; no upstream-derived vectors for IQ1_S exist.
• F5: cpu_quant_dot.cpp:591 claims this kernel reads bsums "and the others do not"; Q2_K, Q4_K and Q5_K already do.
• F6: cpu_quant_iq_tables.h:1-5 still declares the file a 1:1 mirror of one repo while carrying a table from another, and indexes neither new table.
• F7: the census tests claim TOTAL coverage of 1702 records but enumerate 6 of 7 encodings (F32 missing), summing to 864.

Remediation

Seal kIq1sDelta against upstream IQ1S_DELTA as the grid is sealed, and commit golden vectors for both encodings so at least one reference is not the tree's own decoder. Set the NMSE ceiling from the measurement or fix the unstable n=1 estimator instead of widening globally. Add the two missing dequant cases.

Owning row: ENG-EXPERT-STREAM.

[localai-bot]

Fixed by #1026 (4f2d91756).

F1 closed properly rather than patched: the implementer BUILT both pinned oracles (ggml-org/llama.cpp @ 237ad9b96 and unslothai/llama.cpp @ 36fe8e1cc) and had each decode real checkpoint bytes through its own ggml_get_type_traits(type)->to_float, then committed those as tests/vt/iq1_golden_vectors.h. Bit-exact on all 1024 values per encoding. The reference is finally external to the tree, which is what F1 was actually about. kIq1sDelta is additionally sealed by value against upstream IQ1S_DELTA.

All three mutations verified UNCAUGHT before and CAUGHT after. The operator re-ran the delta mutation independently: 4 cases / 2050 assertions fail where it previously passed all 490 tests with assertion counts unchanged.

F2: ceiling set from measurement to 6e-4 (was 2e-3). The repair also found something this issue did not: under a doubled delta, iq1_xxxs NMSE moves the WRONG way (3.109e-4 -> 1.420e-4), so no ceiling could ever catch it there. That makes the goldens the primary fix and the ceiling secondary, which is the right ordering.

F3: case 19 and case 66 added, and ggml 18 (IQ3_XXS) fixed too, since it was the same defect class in the same switch and could be gated by the same oracle.

F5/F6/F7: comment corrected, table header now indexes the constants and flags the fork-sourced table, and the census asserts its buckets sum to 1702 (the red-before printed exactly the 864 == 1702 mismatch this issue reported).

Operator gate: 492/492, exit 0.