fix: discover Criterion samples at their real depth, under new/ (#95)

[Xof]

Closes #95 (BENCH-1, BUG/correctness).

Stacked on #129 (fix/94-file-size-stride). Base retargets to main when that merges.

The defect

discover_cells bounded its walk to exactly depth 4:

// Each leaf is at depth 4 (criterion_dir/row/mode/size/sample.json).
for entry in walkdir::WalkDir::new(criterion_dir).min_depth(4).max_depth(4)

Criterion 0.5 writes <row>/<mode>/<size>/new/sample.json — depth 5. On this
repo's real tree:

$ find target/criterion -maxdepth 4 -name sample.json   # 0 hits
$ find target/criterion -maxdepth 5 -name sample.json   # 2 hits

So the walk matched nothing, and it failed silently: NoCellsFound never
fires because aux_metrics.jsonl still supplies cells via the Step-3 leftover
path. Every micro-grid cell reached the renderers with timing: None, and
every p50/p95/p99 in summary.md rendered as an em-dash.

The one test covering this walked a fixture tree that omits the new/ level,
so the suite passed against a layout Criterion never produces.

The fix

Walk depth 5, and require the leaf's parent directory to be named new.

The new check is the load-bearing half, not the depth change: Criterion
retains the previous run under a sibling base/, which shares the
(row, mode, size) key and would otherwise emit a duplicate Cell.

Tests

The fixture tree gains the new/ level (so it matches what Criterion actually
writes) plus a base/ sibling holding times 10x larger. Two things now fail if
base/ leaks in: an explicit exactly-one-cell count, and the pre-existing
p50 == 3000 assertions (base/ would give 30000).

Counterfactual — reverting the walk to depth 4, all three discover tests fail:

test summary::discover::tests::discover_cells_joins_criterion_with_aux ... FAILED
test summary::discover::tests::discover_cells_handles_missing_aux_gracefully ... FAILED
test summary::discover::tests::discover_cells_handles_malformed_sample_json ... FAILED

copy_raw_archive was always unbounded-depth and already archived the new/
path correctly — only summarize_smoke's expected archive path needed updating
to match the fixture move.

End-to-end verification

Against the real target/criterion, summarize now reports:

smoke/chisel-strict/256B -> p50 84481021.0 p95 89508941.85

Before the fix that cell was emitted with no timing at all.

Verification

cargo test all green, cargo clippy --all-targets -- -D warnings clean,
cargo fmt --check clean.