[ronleizrowice-ant]

Two ways past the check "one wrap, one configuration". Both came with the walk in turns (#628), which is on main and in no release.

What was wrong

A wrap starts over when it finds another configuration object than it started with.

• The working directory is not covered. A relative entry (. as a write root) is resolved against it wherever it is read. A host that calls chdir() in a turn of the walk gets a command put together from both directories: denies walked from the old one, . bound back from the new one. Where the new one lies inside a folder a ** rule denies, what the rule matches there comes back readable.
• An object can come again. initialize() keeps the caller's own object, so reset() and initialize() with it again, during a wrap, looks like no change. A file written before that is not denied, against linux: a wrap that starts over reads the disk again #644's invariant.

What changes

One predicate in place of the identity test, in the walk's steps and after the last await. A wrap also starts over when:

• a configuration has been installed since it began: a count, raised by initialize() and updateConfig();
• process.cwd() has changed (5 to 15 ns a call on Node 22 and Bun 1.4). A directory that was removed counts as a change.

The Linux wrapper awaits once more after that, for its scan for dangerous names, and resolves relative entries on both sides of it. It notes the working directory at its start and throws WorkingDirectoryChanged right after the scan if that has changed, before anything is left on the host; the manager starts over. (Second commit.)

Each new term only makes a wrap start over more often. The bound after two restarts is as it was for the walk; the scan still gives the thread back.

Known limit

Mutating the object handed to initialize() is still not seen. Cloning it there would lose its function-valued fields.

Measured

Real bubblewrap, a walk over 400,150 files, 3 runs a case:

during the walk	main	here
chdir() into a folder a ** rule denies, allowWrite: ['.']	the file is read	refused
a file is written, then reset() and initialize(sameObject)	the file is read	refused
chdir() during the scan (a stand-in that takes 0.4 s), allowWrite: ['.']	the old directory is the write root	the new one is; the scan ran twice

Six rules over the same tree, undisturbed: 1.0 s on both.

Testing

Four new cases: chdir() in the middle of the walk, in its last listing and during the scan, and the second install. All fail on main's source. Six mutations: five fail a test. The sixth, dropping the object test, survives: every install counts now, and the test stays for the paths that set the configuration to undefined.

[antdres]

/stamp