FileStructureWalker (pass HgFile, HgFolder to callable; which can ask for VCS data from any file)
External uses: user browses files, selects one and asks for its history 
Params: tip/revision; 
Implementation: manifest

Log --rev
Log <file>
HgDataFile.history() or Changelog.history(file)?


Changelog.all() to return list with placeholder, not-parsed elements (i.e. read only compressedLen field and skip to next record), so that
total number of elements in the list is correct

hg cat
Implementation: logic to find file by name in the repository is the same with Log and other commands


Revlog
What happens when big entry is added to a file - when it detects it can't longer fit into .i and needs .d? Inline flag and .i format changes?

What's hg natural way to see nodeids of specific files (i.e. when I do 'hg --debug manifest -r 11' and see nodeid of some file, and 
then would like to see what changeset this file came from)?

----------
+ support patch from baseRev + few deltas (although done in a way patches are applied one by one instead of accumulated)
+ command-line samples (-R, filenames) (Log & Cat) to show on any repo
+buildfile + run samples
*input stream impl + lifecycle. Step forward with FileChannel and ByteBuffer, although questionable accomplishment (looks bit complicated, cumbersome)
+ dirstate.mtime
+calculate sha1 digest for file to see I can deal with nodeid. +Do this correctly (smaller nodeid - first)
*.hgignored processing
+Nodeid to keep 20 bytes always, Revlog.Inspector to get nodeid array of meaningful data exact size (nor heading 00 bytes, nor 12 extra bytes from the spec)
+DataAccess - implement memory mapped files, 
+Changeset to get index (local revision number)
+RevisionWalker (on manifest) and WorkingCopyWalker (io.File) talking to ? and/or dirstate (StatusCollector and WCSC) 
+RevlogStream - Inflater. Perhaps, InflaterStream instead? branch:wrap-data-access
+repo.status - use same collector class twice, difference as external code. add external walker that keeps collected maps and use it in Log operation to give files+,files-  
+ strip \1\n metadata out from RevlogStream
+ hash/digest long names for fncache
+Strip off metadata from beg of the stream - DataAccess (with rebase/moveBaseOffset(int)) would be handy
+ hg status, compare revision and local file with kw expansion and eol extension

write code to convert inlined revlog to .i and .d

delta merge
DataAccess - collect debug info (buffer misses, file size/total read operations) to find out better strategy to buffer size detection. Compare performance.
RevlogStream - inflater buffer (and other buffers) size may be too small for repositories out there (i.e. inflater buffer of 512 bytes for 200k revision)  


Parameterize StatusCollector to produce copy only when needed. And HgDataFile.metadata perhaps should be moved to cacheable place?

Status operation from GUI - guess, usually on a file/subfolder, hence API should allow for starting path (unlike cmdline, seems useless to implement include/exclide patterns - GUI users hardly enter them, ever)
  -> recently introduced FileWalker may perhaps help solving this (if starts walking from selected folder) for status op against WorkingDir?

? Can I use fncache (names from it - perhaps, would help for Mac issues Alex mentioned)
? Does fncache lists both .i and .d (iow, is it true hashed <long name>.d is different from hashed <long name>.i)
 
??? encodings of fncache, .hgignore, dirstate
??? http://mercurial.selenic.com/wiki/Manifest says "Multiple changesets may refer to the same manifest revision". To me, each changeset 
changes repository, hence manifest should update nodeids of the files it lists, effectively creating new manifest revision.

? subrepos in log, status (-S) and manifest commands

? when p1 == -1, and p2 != -1, does HgStatusCollector.change() give correct result?

Commands to get CommandContext where they may share various caches (e.g. StatusCollector)
Perhaps, abstract classes for all Inspectors (i.e. StatusCollector.Inspector) for users to use as base classes to protect from change?

-cancellation and progress support
-timestamp check for revlog to recognize external changes
-HgDate or any other better access to time info
-(low) RepositoryComparator#calculateMissingBranches may query branches for the same head more than once 
	(when there are few heads that end up with common nodes). e.g hg4j revision 7 against remote hg4j revision 206 

>>>> Effective file read/data access
ReadOperation, Revlog does: repo.getFileSystem().run(this.file, new ReadOperation(), long start=0, long end = -1)
ReadOperation gets buffer (of whatever size, as decided by FS impl), parses it and then  reports if needs more data.
This helps to ensure streams are closed after reading, allows caching (if the same file (or LRU) is read few times in sequence)
and allows buffer management (i.e. reuse. Single buffer for all reads). 
Scheduling multiple operations (in future, to deal with writes - single queue for FS operations - no locks?)

WRITE: Need to register instances that cache files (e.g. dirstate or .hgignore) to FS notifier, so that cache may get cleared if the file changes (i.e. WriteOperation touches it).  

File access:
* NIO and mapped files - should be fast. Although seems to give less control on mem usage. 
* Regular InputStreams and chunked stream on top - allocate List<byte[]>, each (but last) chunk of fixed size (depending on initial file size) 


* API
  + rename in .core Cset -> HgChangeset,
  + rename in .repo Changeset to HgChangelog.Changeset, Changeset.Inspector -> HgChangelog.Inspector
  - CommandContext
  - Data access - not bytes, but ByteChannel
  - HgRepository constants (TIP, BAD, WC) to HgRevisions enum
  - RevisionMap to replace TreeMap<Integer, ?>
  + .core.* rename to Hg*
  + RepositoryTreeWalker to ManifestCommand to match other command classes 

* defects
  + ConfigFile to strip comments from values (#)

<<<<<
Performance.
after pooling/caching in HgStatusCollector and HgChangeset
hg log --debug -r 0:5000 and same via Log/HgLogCommand: approx. 220 seconds vs 279 seconds. Mem. cons. 20 vs 80 mb.
after further changes in HgStatusCollector (to read ahead 5 elements, 50 max cache, fixed bug with -1) - hg4j dumps 5000 in
93 seconds, memory consumption about 50-56 Mb

IndexEntry(int offset, int baseRevision) got replaced with int[] arrays (offsets - optional)
for 69338 revisions from cpython repo 1109408 bytes reduced to 277368 bytes with the new int[] version.
I.e. total for changelog+manifest is 1,5 Mb+ gain   

ParentWalker got arrays (Nodeid[] and int[]) instead of HashMap/LinkedHashSet. This change saves, per revision:
was: LinkedHashSet$Entry:32 + HashMap$Entry:24 + HashMap.entries[]:4 (in fact, up to 8, given entries size is power of 2, and 69000+ 
	elements in cpython test repo resulted in entries[131072].
	total: (2 HashMaps) 32+(24+4)*2 = 88 bytes
now: Nodeid[]:4 , int[]:4 bytes per entry. arrays of exact revlog size
	total: (4 Nodeid[], 1 int[]) 4*4 + 4 = 20 bytes
for cpython test repo with 69338 revisions, 1 387 224 instead of 4 931 512 bytes. Mem usage (TaskManager) ~50 Mb when 10000 revs read
	
<<<<<

Tests:
DataAccess - readBytes(length > memBufferSize, length*2 > memBufferSize) - to check impl is capable to read huge chunks of data, regardless of own buffer size

ExecHelper('cmd', OutputParser()).run(). StatusOutputParser, LogOutputParser extends OutputParser. construct java result similar to that of cmd, compare results

Need better MethodRule than ErrorCollector for tests run as java app (to print not only MultipleFailureException, but distinct errors)
Also consider using ExternalResource and TemporaryFolder rules. 


=================
Naming:
nodeid: revision
int:	revisionIndex (alternatives: revisionNumber, localRevisionNumber)
BUT, if class name bears Revision, may use 'index' and 'nodeid'
NOT nodeid because although fileNodeid and changesetNodeid are ok (less to my likening than fileRevision, however), it's not clear how
to name integer counterpart, just 'index' is unclear, need to denote nodeid and index are related. 'nodeidIndex' would be odd.
Unfortunately, Revision would be a nice name for a class <int, Nodeid>. As long as I don't want to keep methods to access int/nodeid separately
and not to stick to Revision struct only (to avoid massive instances of Revision<int,Nodeid> when only one is sufficient), I'll need to name
these separate methods anyway. Present opinion is that I don't need the object right now (will have to live with RevisionObject or RevisionDescriptor
once change my mind) 

Handlers (HgStatusHandler, HgManifestHandler, HgChangesetHandler, HgChangesetTreeHandler)
methods DO NOT throw CancelledException. cancellation is separate from processing logic. handlers can implements CancelSupport to become a source of cancellation, if necessary
methods DO throw HgCallbackTargetException to propagate own errors/exceptions
methods are supposed to silently pass HgRuntimeExceptions (although callback implementers may decide to wrap them into HgCallbackTargetException)
descriptive names for the methods, whenever possible (not bare #next)