Measure and record what pre-encoding costs in time, peak memory and disk

## Scope

Measure what the assisted Q2_0 pre-encoding stage costs to run, and record it where a reader planning a pack can find it: elapsed time, peak resident memory, and peak disk for the temporary mixed GGUF.

## Why this is its own ticket

The encoder shipped in #604 without this. An earlier revision of that work carried a `PreEncodeCost` record, peak-RSS measurement in the encoder program, an extra `model_packed` field and a console report. It was removed during review as unrequired scope: [ADR-0032](https://github.com/Alberto-Codes/vramfit/blob/main/docs/adr/0032-assisted-q2-encoder-home.md) decision 3 names only the encoder revision and the matrix provenance, and the removed component justified itself against that record's **open questions** rather than its decisions.

Removing it was the right call for #604. The measurement is still worth having, which is why it lands here instead of being dropped.

## What this discharges

The Acceptance bullet on #601 — "The preprocessor's peak memory and disk are measured and recorded" — is **not met** by #604 and is deferred to this ticket.

## Notes for whoever takes it

- The cost is a property of a real run at real sizes. A figure from a small fixture is not the answer; the target's row widths are 2688 and 1856.
- Peak disk matters as much as memory: the stage writes a temporary mixed GGUF beside the base, and a reader planning a pack needs to know the headroom required.
- Decide deliberately where the figure belongs. A run-log field, a reference page, or both are all defensible; a record that goes stale silently is not.

Start with the assisted Q2_0 pre-encoding stage and read docs/adr/0032-assisted-q2-encoder-home.md for the related design context. Run it at the target row widths of 2688 and 1856, measure elapsed time, peak resident memory, and peak disk for the temporary mixed GGUF, then record the results where pack planning can find them without silently becoming stale.

Written by the indexing model from the issue text.