`compress_model` takes per-tensor candidate `Chain`s, runs each forward DAG via [`forward_chain`], routes every terminal plane through the capability-keyed [`dispatch`] menu, fits Pass-2 shared state on the chosen terminals, and writes a chain-registry-backed `.ptwm` container.

Returns the per-tensor chosen chain *registry index* (after dedup) for audit. The order matches the input slice. **Pass-2 ordering note.** Shared-state fitting is performed *after* Pass-1 multi-chain selection. This means the multi-chain ranker sees only inline trial-encode payloads. For monotone shared-state codecs (PGC, O1SAC), Inline is the lower bound on what the chain can achieve, so the multi-chain ranking remains correct; per-terminal final encoding may later switch to a `Shared` variant if it wins the re-trial.

`element_width` and `is_nibble_packed` come from `transforms::source`, which owns the dtype-code mapping. This function used to inline its own copy of that match. The copy drifted: it kept packed FP4 in the byte catch-all, so every plane reaching the trial encode was byte-width and not nibble-packed, and `PerGroupCodebook` could accept none of them. Call the mapping, do not restate it. `length_bytes` is the raw byte count; `Layout::Rows{row_len}` applies when `shape` is supplied, using the row-major last dimension. Public so the PyO3 binding (`ptwm-py`) can build the same source descriptor without duplicating the dtype-code → element-width mapping.