fit() aliases caller-supplied fixed_W/fixed_H_expr/fixed_H_im arrays instead of copying (crash on read-only input, latent aliasing)

[mkb-phoenix]

Where: cellpie_main.py, intNMF.fit(), lines 190, 200, 212.

def fit(self, adata, fixed_H_expr=None, fixed_H_im=None, fixed_W=None):
    ...
    if fixed_W is None:
        self.theta, theta_it = self._HALS_W(...)
    else:
        self.theta = fixed_W          # line 190 — direct reference, no copy
        theta_it = 0
    ...
    if fixed_H_expr is not None:
        self.phi_expr = fixed_H_expr  # line 200 — same pattern
        self.phi_expr_it = 0
    ...
    if fixed_H_im is not None:
        self.phi_im = fixed_H_im      # line 212 — same pattern
        self.phi_im_it = 0

Symptom: fit() ends with in-place thresholding on its own factor attributes (self.phi_expr[self.phi_expr < 1e-10] = 0, etc). When a caller passes an already-fitted factor matrix in via fixed_H_expr/fixed_H_im/fixed_W — the documented mechanism for holding one factor fixed while fitting another sub-block, which is exactly how this package's own model_selection() bi-cross-validation is implemented internally — that in-place write can crash with ValueError: assignment destination is read-only, if the supplied array happens to be a read-only view (we hit this from a DataFrame.values extraction on one pandas version).

Root cause: self.theta = fixed_W etc. binds the attribute directly to whatever object the caller passed in, rather than copying it. Two distinct problems follow: (1) whether the crash above triggers depends on whether the caller's array happens to be writable, which is version/call-path dependent and not guaranteed either way by fit()'s own signature; (2) even when it doesn't crash, two separate intNMF instances that are each given the same array via fixed_* now alias the same underlying memory — one instance's later in-place threshold step would silently mutate data the other instance still holds a reference to.

Suggested fix: copy on assignment.

self.theta     = np.array(fixed_W,      copy=True)
self.phi_expr  = np.array(fixed_H_expr, copy=True)
self.phi_im    = np.array(fixed_H_im,   copy=True)

Scope note: _initialize_nmf() (lines 474-655) has its own, separate handling of the same fixed_W/fixed_H_expr/fixed_H_im parameters, which we have not touched or tested — flagging as unconfirmed/out of scope rather than claiming it's also affected.

---

Found while running CellPie's own cp_utils.model_selection() bi-cross-validation at production scale. Happy to share our patched fork if useful (commit pinned at 8880d673c, main branch).