I thought about how to make it harder for developers to forget the encode/decode step in future changes. The approach I ended up with is to couple the conversion with the read/write itself: every place that moves a value between a constituent cursor and the outside world now goes through a small set of wrappers that do the conversion internally. Anyone writing a new read or write path will naturally reach for these wrappers and get the encoding for free. It should also simplify s_tombstone_encoding.py, since the reviewed caller inventories collapse down to the wrappers.

The resulting structure:

- __clayered_deleted_encode() has exactly one caller: __clayered_put(). Storing a value into a constituent is now indivisible from encoding — modify rewrites also go through it, via a new WTI_CLAYERED_PUT_MODIFY_REWRITE op.
- __clayered_deleted_decode() has two callers in the layered cursor logic: __clayered_promote_value() (surfacing a constituent value on the user cursor) and __clayered_modify_apply_decoded() (decoding the base value a modify applies against). The only other caller is __wt_clayered_ingest_to_stable_value(), the drain conversion, which keeps its own contract.

So the cursor code never reads or writes a value except through these named wrappers. The only stores outside __clayered_put() are the raw tombstone delete marker and the plain modify() fast path (FIXME-WT-18057).

Two behavioral notes:

- A modify rewrite of an encoded-namespace value now re-searches instead of using the held cursor position, because __clayered_put() sets the key. This only affects the rare namespace paths; the common plain-modify path is untouched.
- The encode scratch buffer moved into __clayered_put() and is freed before it returns. That is safe because a successful insert or update re-points the constituent's value at internal update memory, and a failed non-overwrite insert returns a cursor-local copy of the existing value.

It's not a small change, so if you don't think it makes things better, I don't want to block this PR on it. Here is the full diff.

[wrappers.patch](https://github.com/user-attachments/files/30579398/wrappers.patch)