What you are asking about is morphological productivity of English suffixes. Productivity in this sense means using morphemes to coin new words. (It just so happens my thesis deals with morphological productivity!)
There are a number of different aspects of the language system that influence productivity and the makeup of the lexicon.
Synchronic change
As a language evolves through time, different suffixes rise in productivity and later fall in productivity. For example (from Anshen & Aronoff 1999), the suffix -ment entered into the language around 1250 according to the OED, through whole-word borrowings from French (many, if not most, suffixes enter a language this way). It converts verbs to nouns. The number of new words coined with -ment steadily rose up until around the 16th century, when it sharply began to decline. The OED cites only 2 new words since 1950 that were coined using -ment. However, once a word has been derived, and has stabilized in the lexicon, it becomes fossilized. That is to say, when we use the word "electricity", we aren't actively calculating electric + ity = electricity
each time we use the word. Another example of a suffix that used to be productive that is no longer productive is -th (as in "warmth", "length", "width"). We don't make new words with this suffix. So, many of the suffixes we see in older words are no longer productive, but the word is still part of our lexicon.
Phonological restrictions
In other cases, two suffixes can be productive, but be restricted by phonology (that is, syllable structure and word stress). For example, the suffixes -ize and -(i)fy both are used to create verbs, but -ify is restricted to one-syllable words, two-syllable words with an unstressed final syllable ending in a vowel (with a few rare exceptions, e.g. "solidify"). For example, "glorify", "testify", "pacify". Of course, -ize is used in the other cases, so we get "revolutionize", "customize", etc. People make this choice when creating new words without even thinking about it – one will just sound "right". This is one of the awesome things about language! Also, referring back to the above paragraph, these phonological restrictions may change over time and one suffix might take over in all new cases. (This comes from Marchard 1969.)
Semantic restrictions
Many times, a suffix is chosen because it carries certain semantic features, like the suffix -er, which converts verbs to nouns, but also adds the meaning "one who does X". So another verb-to-noun suffix might be productive but doesn't carry that meaning.
Pragmatic restrictions
Sometimes, one suffix works better in a certain situation; think of formal/informal distinctions. I might say "naivety" if I wanted to sound very formal or educated, or I might say "naiveness" in a less formal situation. Another example is that we have the word "neurological", but it is my understanding that neuroscientists often use the word "neurologic" – the alternate word distinguishes the sense of the layman's term from the nuances of the field-specific term.
Sometimes it is unpredictable
Sometimes, a word+suffix combo gets fossilized with a certain meaning and the same word with a different suffix gets fossilized with a different meaning. Think "historical" (happened in the past) vs. "historic" (extra meaning of having a great impact on history -- can be said about present-day events). It's not -ic or -ical that is adding this extra meaning, it's that the words have fossilized in the language and then developed different nuances.
Addressing some of your examples
The word excessive was first cited at the end of the 14th century, so this is a fairly old word. The -ive suffix came from French originally. This suffix distinguishes itself from others on semantic grounds, as explained in the OED (emphasis mine):
the suffix is largely used in the
modern Romanic languages, and in English, to
adapt Latin words in -īvus, or form words
on Latin analogies, with the sense
‘having a tendency to, having the
nature, character, or quality of,
given to (some action)’. The meaning
differs from that of participial adjs. in
-ing, -ant, -ent, in implying a permanent or habitual quality or
tendency: cf. acting, active,
attracting, attractive, coherent,
cohesive, consequent, consecutive.
As for the non-words, excessic/excessal/excessical, the most likely reason they were not created is because excessive was already fossilized in the language and, at least thus far, there has been no special nuance or meaning that has demanded an excessic/excessal form that would set it apart from excessive, or cause it to replace excessive. This disallowance of synonymous coinings is also sometimes called Blocking (Aronoff 1976) or the Avoid Synonymy Principle (Kiparsky 1983). It is something that is still being worked out by morphologists today.
Best Answer
Having mulled this over in my head for a bit, I finally came—through the help of @starplusplus’ comment—to a distinction which I think holds up quite nicely in the vast majority of cases. There will always be odd ones out that are just completely idiomatic and do not hold up to logical scrutiny (many, in fact—this is language we’re dealing with), but the following tallies with both my instinctive, hard-to-verbalise gut feeling and also nearly all the examples I can come up with.
Briefly:
All of these words are used (non-figuratively) of three-dimensional objects, and they both tell us something about two of these dimensions: they tell us nothing of length, but they do say something about width and depth.
Objects that do not have a particular, notional ‘surface’
When you use thin/thick, you are giving no particular importance to width and depth in relation to each other—their values are implied to be equally distributed, or any difference in their values is considered unimportant. (Though see below)
When you use narrow/broad, on the other hand, you are ascribing particular importance to one dimension (width) over the other (depth); or, in some cases, you are describing that one dimension (width) in particular has a higher or lower value that normally seen. Depth is either not considered relevant, or is considered relevant only as the value against which the narrow/broad thing is compared.
Thus for example, neither narrow ribbon nor thin ribbon tells us anything at all about the length of the ribbon in question. But where thin ribbon tells us only that the width and depth dimensions have small values—leaving aside as unimportant whether they are in fact identical values or not—narrow ribbon tells us nothing much about the depth dimension of the ribbon, but only the width dimension, which is said to have a smaller value than normally seen (in the archetypal, prototypical ribbon that of course looks different in everyone’s mind).
Objects that have a particular, notional ‘surface’
If something is of such a shape that we consider one of its surface areas to be the surface (such as tables, mattresses, or perhaps even ribbons), thin/thick is often used almost contrastively to narrow/broad in that it then emphasises the depth dimension in the same way that narrow/broad emphasises the width dimension. I would say that this is a narrowing (!) down of the two-dimensional meaning to one dimension, for the simple reason that we just do not talk about the dimensions of objects that we consider to have a single ‘important’ surface with indiscriminate regard to the relationship between the dimensions.
Consider a mattress as an example. It has a surface, which is always the side that faces up on our beds. We may wish to describe the length, width, or depth of the mattress individually, but there is no practical need to ever describe collectively the width and depth of it with no real distinction between the two.
Consider then instead a wooden beam. We don’t generally consider beams to have just one surface—all four sides (excluding the ends) are equally important, because no one of them has a particular, inherent usage. We may still wish to address all three dimensions individually, in which case we still use narrow/broad to refer to the width dimension, and thin/thick to refer to the depth dimension; but if there is no implied or explicit comparison with the width dimension, then thin/thick can still quite easily be used in its more generic, two-dimensional sense, to indicate that both the width and depth dimensions of the beam have relatively high values, regardless of what the exact relationship between these values are.
(Incidentally, in this view, I don’t consider narrow waist to be so irrational after all: you are simply commenting specifically on the width dimension, which is of course the diameter from the left side of the body to the right; i.e., someone with a narrow waist looks narrow from the front, but not necessarily from the side. Thin waist would instead refer to someone who looks thin from all sides.)