Encoder: brute-force config search per-image (zenjpeg/coefficient pattern)

@lilith is already working on this.

Since Apr 26, 2026.

**Future work** — discussed during the libwebp-parity sprint (audit branch `audit/libwebp-parity-2026-04-26`).

## What

Add a "config search" mode to the encoder, modeled after the brute-force / Pareto-search pass that `coefficient` and `zenjpeg` already do. For each input image, the encoder probes multiple (preset, sns_strength, filter_strength, segments, partition_limit, cost_model, ...) tuples and picks the one that wins on the user-chosen objective (smallest size at target quality, best butteraugli at target size, etc.).

## Why

Even with good defaults, individual images respond differently to perceptual tuning (e.g., gradients vs. text-heavy graphics vs. photos). A per-image search routinely produces ~5–10% better RD curves on `zenjpeg`'s benchmark corpus.

## Sketch

- New API: `EncoderConfig::with_search(SearchObjective::MinSize { target_quality }, max_probes: 16)`.
- Internal probe loop: encodes the image at each candidate config, decodes for quality (zensim / fast-ssim2 / butteraugli), records (size, quality), filters by objective, returns the best.
- Pareto-frontier extraction so the user can pick along the curve.
- Honors `CostModel` selection: if user picks `StrictLibwebpParity`, only libwebp-equivalent configs are probed.
- Caches per-image features (DCT histogram, edge density, color count) so the probe set can be pruned heuristically (e.g. low-color images skip Photo preset).

## References

- `coefficient` crate's per-codec config sweep loop (in `~/work/coefficient/`)
- `zenjpeg`'s adaptive-quantization pre-pass

## Dependencies

- Bit-exact libwebp parity mode (issue tracking the StrictLibwebpParity north star) — needed as the comparison reference and as one of the candidate configs.
- A solid `parity_baseline`-style harness for size+quality+speed measurement (`dev/parity_baseline.rs` already exists).

## Status

Backlog. No implementation in this session.