No way to see what changed between two mirrors

[xroche]

--update already works out which resources are new, which changed and which the server reports unchanged, then drops it: the flags reach file_notify() and go no further than a log line, and deletions exist only as a side effect of purging. --changes (-%d) keeps all of it and writes hts-changes.json next to the log, with the four lists, their counts and a one-line summary in the log.

"Changed" means the bytes differ, not that the server re-sent the resource. Comparing the mirrored files does not work for that: HTTrack stamps every parsed page with the crawl date through the footer, so those bytes move on every run and the whole site reads as changed. The comparison is on payloads instead. For parsed pages the previous payload is the body the cache kept; everywhere else the mirrored file is the payload verbatim, and the previous copy is hashed just before it gets overwritten. With no cache and nothing to hash, parsed pages fall back to the 200-versus-304 signal. The cache also carries the mirror's file index, so --cache=0 leaves no previous mirror to subtract from: nothing is reported gone, and first_crawl comes out null rather than guessed.

Accounting is keyed on the mirror-relative path, not the URL, so a redirect and its target that share a save name are one entry, and only the first notify for a file samples its pre-run state, so a retried transfer is not counted twice. What counts as already mirrored comes from the previous run's file index rather than from the file being on disk: a partial left by this crawl's own failed attempt is on disk but was never part of the previous mirror. A resource the crawl tried and failed to fetch drops out of that index too, but its copy is untouched, so it is reported unchanged rather than gone. The deleted set is now computed whether or not purging is on; unlinking still happens only under --purge-old.

The accumulator is locked on both sides. FTP transfers notify from a thread the crawl never joins, so one can arrive while the report is being resolved and written, and the report path used to run unlocked over an array realloct() was moving underneath it. Writing the report also seals the accumulator: a notify that lands afterwards is dropped instead of starting a second report nobody will write. With the feature off the hooks return before doing any work, so an --update without --changes performs the same cache reads as master and produces the same mirror.

htsopt.h gains two tail-appended fields on httrackp (changes, changes_state), so this is an installed-header change and anything embedding the struct needs a rebuild. I have not touched VERSION_INFO. Content diffs and keeping the previous copy of a changed page are out of scope on purpose: both change what a mirror directory holds and run into the purge invariant.

tests/93_local-changes.test is a three-pass crawl whose routes all answer 200 with no Last-Modified and no ETag, so a naive implementation calls the whole site changed; it asserts each bucket exactly, covers a redirect, a retried direct-to-disk transfer, a transfer that never completes, a pass with purging on and a mirror crawled with the cache off. httrack -#test=changes covers the bucket rules and the JSON escaping, invalid UTF-8 included, and -#test=changes-race runs notifier threads against the report path for the FTP shape (clean under ThreadSanitizer, fifteen races with the lock removed). Every rule above was graded against a build with that rule mutated, to check the tests go red. The file format is documented in html/changes.html, and the WebHTTrack toggle, man page and command-line guide are updated; WinHTTrack and Android only need the flag added to the argv they already build.

Closes #714