get_record_stream/ContentFactory should expose compression parents

For instance there is this code in knit.py, which probably should instead be using an explicit compression parent.

            # Add any records whose basis parent is now available.
            added_keys = [record.key]
            while added_keys:
                key = added_keys.pop(0)
                if key in buffered_index_entries:
                    index_entries = buffered_index_entries[key]
                    self._index.add_records(index_entries)
                    added_keys.extend(
                        [index_entry[0] for index_entry in index_entries])
                    del buffered_index_entries[key]

I'm not sure that Martin copy & pasted the right code snippet. I would have thought more of:

                index_entry = (record.key, options, access_memo, parents)
                buffered = False
                if 'fulltext' not in options:
                    basis_parent = parents[0]
...
                    if basis_parent not in self.get_parent_map([basis_parent]):
                        pending = buffered_index_entries.setdefault(
                            basis_parent, [])
                        pending.append(index_entry)
                        buffered = True
                if not buffered:
                    self._index.add_records([index_entry])

Actually the real problem is down in _KnitGraphindex.add_records, in that code you have:

                if self._deltas:
                    if 'line-delta' in options:
                        node_refs = (parents, (parents[0],))
                    else:
                        node_refs = (parents, ())

So when this delta is added to the index, it implicitly assumes that parents[0] is the compression parent for this delta.

I'll also note that for knit repositories, they cannot represent a delta which is not against their left-hand ancestor. Only Pack indexes have a field for compression parent. So some care is needed, in that pulling a delta into a repository that cannot represent it will need to be extracted up to a full text and then delta compressed again.

So it was worked out that Knit*Factory objects just conform to the api that compression parents are always the left-hand parent. And other factory objects (so far) don't have compression parents, so it didn't matter.