fix(io): fail fast on unsupported GGUF tensor types; load Q4_0/Q5_0/Q5_1 packed (#919)

[michalharakal]

What

StreamingGgufParametersLoader silently skipped any tensor type outside its when — a SKIP progress string was the only signal, the load "succeeded" with missing weights, and the failure surfaced far away in graph build / the first forward pass (the load-time half of #654). The deprecated legacy GgufParametersLoader already hard-failed here; the streaming loader regressed that. Full analysis in #919.

Two distinct gaps hid in that else:

1. Truly unsupported types (e.g. Q4_1) → now an eager pre-scan of the tensor directory throws IllegalArgumentException before any tensor is delivered, naming every offending tensor, its type (raw on-disk value for unknown types, truncated listing past 8 offenders), and the supported set:

   GGUF contains 1 tensor(s) with quantization types this loader does not support: 'blk.0.ffn_down.weight' (Q4_1). Supported types: F32, I32, F16, BF16, Q4_0, Q5_0, Q5_1, Q4_K, Q5_K, Q6_K, Q8_0. Re-quantize the model to a supported format (e.g. Q8_0, Q4_0, Q4_K or F16).

2. Q4_0 / Q5_0 / Q5_1 had packed TensorData (fromRawBytes) and matmul kernels but were missing from the loader's when — a Q4_0 GGUF (a first-class kernel format) loaded with missing weights. They now load as packed blocks, mirroring the Q8_0/Q4_K branches.

Design notes

• Both the pre-scan and the when anchor on a new SUPPORTED_TENSOR_TYPES companion set; the per-tensor else is now a hard IllegalStateException that names the drift if the two ever disagree.
• Dtype-policy mismatches keep their existing skip-via-null behavior (e.g. an I32 tensor under an FP32 load) — that path is a filter consumers rely on, not an error. Only tensor types fail fast, matching the "fail before execution" rule validatePolicy/withPolicy already applies (its kdoc previously acknowledged this exact gap).
• Behavior change: a file that previously "loaded" with skipped tensors now fails at load. CHANGELOG entry included.

Tests

• commonTest StreamingGgufParametersLoaderFailFastTest (5 cases, runs on every target): supported set passes, Q4_1 rejected with tensor name + supported set, unknown raw type value reported verbatim, long offender lists truncated with a count, Q4_0/Q5_0/Q5_1 pinned in the supported set.
• jvmTest: synthesized-GGUF e2e — Q4_1 file throws before any onTensorLoaded callback; multiple offenders all named; Q4_0/Q5_0/Q5_1 load as PackedBlockStorage. The existing GGUF-builder helper was parameterized to cover arbitrary tensor sets.

Verified

• :skainet-io:skainet-io-gguf:jvmTest — 19/19 green (6 loader + 5 fail-fast + 8 policy)
• :skainet-io:skainet-io-gguf:linuxX64Test — green (commonMain + commonTest on Kotlin/Native)
• :skainet-io:skainet-io-core:jvmTest — green (downstream GGUF fixture loads unaffected)
• No other in-repo consumers of the loader

Closes #919

🤖 Generated with Claude Code