Unreliable behavior of hash tables with symbols as keys (regression wrt 4.7.4)

[Ivan Raikov]

The nemo program relies heavily on performing transformations over hash tables with symbols as keys. It works fine with Chicken 4.7.4, but unfortunately there seems to have been a regression in 4.8.0rc1, which results in hash-table-exists? and hash-table-ref failing to find any of the existing keys in the table after a number of transformations have been performed.
The following code reproduces the issue:

(use srfi-1 srfi-69)

(define t (make-hash-table hash: symbol-hash))

(hash-table-set! t 'k1 1)
(hash-table-ref t 'k1)

(for-each
  (lambda (k v)
    (printf "hash-table-set! ~A ~A~%" k v)
    (hash-table-set! t k v))
  (list-tabulate 10 (lambda (i) (gensym 'k)))
  (list-tabulate 10 (lambda (i) i))
  )

(print (hash-table-ref t 'k15))

Perhaps the hash tables stop working after some garbage collections have been performed. This issue is not entirely new; nemo previously used the environments egg, which was also exhibiting similar unreliable behavior, so I had resorted to using strings instead of symbols as keys. Any advice on how to debug this would be appreciated.