[Quchaosheng]

Two fixes for #1282 and #1281.

#1282 - #1266 dropped a document that failed to load, but the rollback
stopped at the document boundary. If main.xml includes sub.xml and then
fails, sub.xml has already registered its trees and they stay behind;
sub.xml also becomes the first document, so instantiateTree() reads
main_tree_to_execute from it. Take a snapshot of
tree_roots/tree_sources/suffix_count before the call and restore it on
failure. That also puts back a tree the call redefined, whose element lives in
the document about to be erased.

#1281 - Tree::~Tree() calls haltTree(), and halting may throw: a
TryCatch with catch_on_halt="true" ticks its catch child, and a tick is
allowed to throw. An exception leaving a destructor calls std::terminate, so
destroying such a tree aborted the process. Mark the destructor noexcept and
swallow the exception.

Both come with tests that fail on the current master:

• XMLParserTest.RejectedIncludeKeepsNoTree and
  RejectedInlineIncludesKeepNoTree see the included tree survive a rejected load.
• TreeDestructorTest.HaltThatThrowsDoesNotTerminate aborts with
  terminate called after throwing an instance of 'BT::NodeExecutionError'.

Full suite: 684 passed, 2 skipped. The XML side is also clean under
-fsanitize=address.

The noexcept on Tree::~Tree() changes a public header. If you would rather
keep the signature alone, dropping the keyword and keeping the try/catch in the
body is enough - say so and I will change it.

---

Maintainer changes (rebased on master after #1290)

• One rollback point. <include> now calls loadDocImpl directly, and only the top-level load call snapshots and restores. On failure it drops every document the call opened (erase(doc_it, end())). Before, a valid included document stayed behind as the "first document" whose main_tree_to_execute a later load picked up. The snapshot also covers subtree_models and current_path, as listed in XMLParser: a rejected file's valid includes stay registered #1282.
• ~Tree(). The noexcept is gone (destructors are implicitly noexcept). A throwing halt is now reported on stderr, as FinallyNode does, instead of being swallowed silently.
• Tests.
  • The temp-dir helper is shared through tests/test_helper.hpp, which removes the getpid call that failed on MSVC.
  • The destructor test moved next to the other catch_on_halt tests and checks that the catch child ran.
  • New RejectedIncludeIsNotTheFirstDocument test.
  • The redefinition test now goes through a valid include of a rejected file; the previous version passed on master.

Behavior change: after a rejected load, trees from its valid includes are no longer registered. A load either succeeds completely or leaves the parser as it was.

Follow-up, not in this PR: a node that throws while halting still stops haltTree() before it reaches the remaining nodes. The PR prevents std::terminate, but sibling nodes may miss onHalted().

[facontidavide]

let me ask this.

Is this a REAL problem YOU have in PRACTICE, or something reported by AI?

I don't have any problem with AI, but I feel like this is solving a problem nobody has

[Quchaosheng]

let me ask this.

Is this a REAL problem YOU have in PRACTICE, or something reported by AI?

I don't have any problem with AI, but I feel like this is solving a problem nobody has

I used DeepSeek as a tool, but the analysis is my own