SLOT 43 DECODED (partially) -- it is a REPORTING request, not an image source.

Length: exactly 10 words (0x65618 SUBK 0xa ; JREQ ; else BadLength 0x10).

Payload words actually read (bit offsets -> payload byte / word index):
  0x65624 *A9(0x40)  byte 8   -> payload word 1   (id #2)
  0x65628 *A9(0x20)  byte 4   -> payload word 0   (id #1)
  0x65650 *A9(0xa0)  byte 20  -> payload word 4   (ROI field)
  0x65668 *A9(0xc0)  byte 24  -> payload word 5   (ROI field)
  0x65674 fallback width  = *(*A13(0x1c0))(0x80)
  0x65678 fallback height = *(*A13(0x1c0))(0xa0)
so the shape is [id1, id2, ?, ?, roiA, roiB, ...] with two live resources
looked up through 0x7EA6FA0 at 0x65640.

The "descriptor tables with 3.0 / 438.0" are NOT descriptors: they are arrays of
IEEE doubles and each MOVI points 8 bytes further into the same array:

  3F533740 -> 00 00 00 00 00 00 08 40  = 3.0
  3F533780 -> next double (8 bytes later)
  3F5337C0, 3F533800, 3F533840, 3F533880 ...   successive doubles

and each load is followed by two `MOVE *A8+,*A14+` (copy 8 bytes = one double)
then `CALLA 0x07FC1CE6 / 0x07FC18F6 / 0x07FC1CDE` -- the printf family (the same
helpers used by the export timing print and by 0x7F638C2, the logger).

So slot 43 formats and prints doubles; it is a diagnostic/report request.

TESTED LIVE (valid c.new_id ids, no-reply model). 17 payload shapes, including
doubles (0x40080000 = high half of 3.0) in every position, single-id and two-id
forms, and ROI-shaped values:

  all-zero, wN=3.0hi for N=1..5, w1..w3, all=3.0hi
  [A,B,0,0,64,64,0,0,0]  [A,A,0,0,64,64,0,0,0]  [A,B,0,0,64,64,1,1,0]
  [A,B,0,0,0,0,0,0,0]    [A,B,1,1,64,64,0,0,0]  [A,B,64,64,64,64,0,0,0]

  EVERY ONE returned BadValue (2).

Therefore:
 * slot 43's exact ROI/validity preconditions were not reconstructed, but
 * its body is printf-dominated and its constants are printed doubles, and
 * nothing observed connects it to src_base or to an image data plane.

It is therefore not credible as "the image-source element that populates
src_base": there is no evidence it writes any image buffer.

Incidental but useful: c.new_id now yields ids like 0x00302000 / 0x00303000 on
this run, confirming again that the per-connection id base varies (0x0040xxxx,
0x0050xxxx, 0x0030xxxx all seen) -- further validating the earlier id fix.
