bug#14054: completing against a hash-table doesn't work as documented

[Jonas Bernoulli]

`completing-read's doc-strings sais I should read `try-completion's
(and `all-completions's) doc-string for details on COLLECTION.

`try-completion's doc-string sais:

>   If COLLECTION is a hash-table, all the keys that are strings or
>   symbols are the possible completions.

(1) To my this implies that all other keys are simply ignored, as it
    doesn't say that all keys _have to by_ strings or symbols.

    However using e.g. a vector or number as key causes a
    `wrong-type-argument' error in `test-completion'.

(2) It explicity sais that a key can be a symbol.

    However this causes the same error:

    (let ((ht (make-hash-table)))
      (puthash 'key 'val ht)
      (test-completion "input" ht))

I tried using hash-tables created with all three predefined hash-table
tests `eq', `eql', and `equal; none worked.

Note that depending on the "order" of elements in the hash-table no
error is raised by `test-completion' if a string key "before" the first
non-string key matches:

no error

    (let ((ht (make-hash-table)))
      (puthash "first" 'val ht)
      (puthash 'second 'val ht)
      (test-completion "first" ht))

error

    (let ((ht (make-hash-table)))
      (puthash 'first   'val ht)
      (puthash "second" 'val ht)
      (test-completion "second" ht))

I have tried with 23.4, 24.2, 24.3, and 24.3.50; they all behave the
this way.