I would like to: represent the input tarballs as a
commit each (which all get merged together as if by git merge -s
subtree), and for quilt packages, each patch as a commit.  But I want
to avoid (as much as possible) reimplementing the package extraction
algorithm in dpkg-source.

dpkg-source does not currently provide interfaces that look like they
are intended for what I want to do.  And dgit wants to work with old
versions of dpkg, so I don't want to block on getting such interfaces
added (even supposing that a sane interface could be designed, which
is doubtful).

So I intend to do as follows.  (Please hold your nose.)

* dgit will untar each input tarball (other than the Debian tarball).

  This will be done by scanning the .dsc for things whose names look
  like (compressed) tarballs, and using the interfaces provided by
  Dpkg::Compression to get at the tarball.

  Each input tarball unpack will be done separately, and will be
  followed by git-add and git-write tree, to obtain a git tree object
  corresponding to the tarball contents.

  That tree object will be made into a commit object with no parents.
  (The package changelog will be searched for the earliest version
  with the right upstream version component, and the information found
  there used for the commit object's metadata.)

* For `3.0 (quilt), dgit will run
    dpkg-source -x --skip-patches

  git plumbing will be used to make the result into a tree and a
  commit.  The commit will have as parents all the tarballs previous
  mentioned.  The metadata will come from the .dsc and/or the final
  changelog entry.

  dgit will then see if it has a series file.  (dgit already rejects
  packages with distro-specific series files, so we need worry only
  about a single debian/patches/series file.)

  If there is a series file, dgit will read it into memory.  It will
  then iterate over the series file, and each time:
    - write into its playground a series file containing one
      more non-comment non-empty line to previously
    - run dpkg-source --before-build (which will apply that
      additional patch)
    - make git tree and commit objects, using the metadata from
      the relevant patch file to make the commit (if available)
    - each commit object has as a parent the previous commit
      (either the previous commit, or the commit resulting from
      dpkg-source -x)

  After this the series file has been completely rewritten in
  this way, the tree should be identical to the results of
  dpkg-source -x.

* For source formats other than `3.0 (quilt)', dgit will do simply
    dpkg-source -x.

  Again, it will make that into a tree and a commit.  (If this commit
  has no changes in its tree from the first tarball commit. then the
  two are squashed together.)

* As currently, there will be a final no-change-to-the-tree
  pseudomerge commit which stitches the package into the relevant dgit
  suite branch; ie something that looks as if it was made with git
  merge -s ours.

* As currently, dgit will take steps so that none of the git trees
  discussed above contain a .pc directory.


This has the following properties:

* Each input tarball is represented by a different commit; in usual
  cases these commits will be the same for every upload of the same
  upstream version.

* For `3.0 (quilt)' each patch's changes to the upstream files appears
  as a single git commit (as is the effect of the debian tarball).
  For `1.0' non-native, the effect of the diff is represented as a
  commit.  So eg `git blame' will show synthetic commits corresponding
  to the correct parts of the input source package.

* It is possible to `git-cherry-pick' etc. commits representing `3.0
  (quilt)' patches.  It is even possible fish out the patch stack as
  git branch and rebase it elsewhere etc., since the patch stack is
  represented as a contiguous series of commits which make only the
  relevant upstream changes.

* Every orig tarball in the source package is decompressed twice, but
  disk space for only one extra copy of its unpacked contents is
  needed.  (The converse would be possible in principle but would be
  very hard to arrange with the current interfaces provided by the
  various tools.)

* No back doors into the innards of dpkg-source (nor changes to
  dpkg-dev) are required.

* dgit does grow a dependency on Dpkg::Compression.

* Knowledge of the source format embedded in dgit is is restricted to
  iterating over tarballs and manipulating debian/patches/series,
  which dgit already does.

* dgit now depends on dpkg-source --before-build idempotently applying
  patches as they successively appear on debian/patches/series.

* Perhaps the git commits generated by dgit to represent patches can
  be made to round-trip nicely into tools like git-dpm and
  git-buildpackage.

  I have found the information about tags in gbp-dch(1), but that
  doesn't seem like it's applicable.

  I have also found the information about tags in gbp-pq(1).  From
  that it looks like I ought to generate "Gbp-Pq: Name" and "Gbp-Pq:
  Topic".

* The scheme I describe avoids introducing a dependency from dgit to
  git-buildpackage.