feat(io): report every weight the loader re-encodes on the way in

[michalharakal]

Closes #1117. Slice 4 of 5 for #1109.

"Why is this model bigger than the file it came from" had no answer except whoever remembered which policy was set. Now each conversion emits a TraceEvent.AdapterInserted naming the tensor and the sizes either side — and a weight that arrives in the form it will be used in emits nothing at all.

Reusing the event rather than adding one

AdapterInserted's own documentation already describes this: "the dispatcher inserted a conversion (dequantize, requantize, gather)". A loader-side re-encode is the same thing at a different moment. Reusing it means the Perfetto, JFR and android.os.Trace exporters need no changes to show it — the third acceptance criterion satisfied by not writing code.

It gained one defaulted field, bytesBefore, plus a derived bytesDelta. A conversion's cost is the difference between two sizes, and one number cannot express it; Perfetto and JFR now carry both.

Three conversions, not one

The issue asked for the conversions the resolver causes, which today means form-driven dequantization alone. But the loader performs two other widenings that nobody ever asked for, and those are the ones most worth seeing:

kind	when	cost
dequantize-on-load	the resolved form says DequantizeTo(FP32)	Q4_K → FP32, ~8×
widen-ternary-no-packed-storage	always, for TQ1_0/TQ2_0 — packed ternary storage does not exist yet (#1033)	~20×, and no flag on this loader turns it off
widen-f16 / widen-bf16	unless keepF16Native / keepBf16Native	2×

Reporting only the conversion someone explicitly requested would answer the easy half of the question. The ternary one in particular is invisible today: it is not a policy, it is a gap, and the trace is where a gap that costs 20× should show up.

Cost when nobody is listening

The sink defaults to NoopTraceSink, and every emission is guarded on isEnabled, so a caller who asked for no trace builds no Formats and allocates nothing.

The tensor is named by parsing the GGUF name as a TensorId — blk.0.attn_q.weight is already dotted and renders back as itself, so no new field was needed to carry a name.

Tests

Five: the silent case is genuinely silent; a dequantized weight reports the right tensor, both sizes and the delta; the ternary widening is caught with its multiple asserted; a narrow float reports its doubling; and the default loader records nothing.

Gate

scripts/pr-gate.sh — all legs passed.

🤖 Generated with Claude Code

[github-actions]

📖 Documentation Preview

The documentation has been built successfully for this PR.

Generated Files:

• Operator documentation: docs/modules/operators/_generated_/
• JSON schema output: operators.json

Artifacts:

• Download the documentation-preview-1123 artifact to view the complete documentation locally.

This comment will be updated automatically when the PR is updated.