feat(core): log_field_fold — reversible log-line columnar transform

[snchimata]

Summary

Adds log_field_fold, a losslessly-reversible columnar transform for templated log lines — the log-line analogue of json_field_fold.

What & why

Most log lines share a fixed skeleton and vary only in a few fields (timestamps, ids, counts). This replaces each variable token (a digit run or 0x… hex literal) with a placeholder to derive a per-line template, then emits each distinct template once plus a compact per-line row of captured values — so the skeleton text is paid for once instead of on every line.

Unlike log_compaction (which only collapses identical adjacent lines, lossy-with-evidence), this compresses distinct-but-templated lines and is fully lossless.

Lossless & safe

• fold_log / unfold_log / round_trips; the pipeline (validate_safety) adopts the fold only when round_trips holds, so it can never lose data — a fold that wouldn't reverse is rolled back rather than emitted.
• A proptest verifies the fold→unfold identity on arbitrary log-shaped input (CRLF, blank lines, and a missing final newline are all preserved byte-for-byte).

Registration

• modes.rs: TransformId::LogFieldFold + ModeEntry — experimental, out of Conservative, plain_text/command_output, ordered before the lossy log_compaction (lossless-before-lossy).
• pipeline.rs: forward dispatch + round-trip safety gate.
• tests/fixtures/mode_matrix.toml: mirror entry.

Ships behind --experimental until its fidelity gate is green (the same path json_field_fold and log_compaction took).

Testing

• All 225 core tests + the full workspace (25 suites) pass, including the proptest round-trip.
• CLI end-to-end: tokenfold compress --experimental on a ~20k-token templated service log → ~34% reduction, losslessly.

Also re-adds the /eval/learned/ gitignore rule for the local research scratch dir.