POC broken alt cancellation 3032

[erights]

Most PRs should close a specific Issue. All PRs should at least reference one or more Issues. Edit and/or delete the following lines as appropriate (note: you don't need both refs and closes for the same one):

Closes: #XXXX
Refs: #XXXX

Description

Add a description of the changes that this PR introduces and the files that are the most critical to review.

Security Considerations

Does this change introduce new assumptions or dependencies that, if violated, could introduce security vulnerabilities? How does this PR change the boundaries between mutually-suspicious components? What new authorities are introduced by this change, perhaps by new API calls?

Scaling Considerations

Does this change require or encourage significant increase in consumption of CPU cycles, RAM, on-chain storage, message exchanges, or other scarce resources? If so, can that be prevented or mitigated?

Documentation Considerations

Give our docs folks some hints about what needs to be described to downstream users. Backwards compatibility: what happens to existing data or deployments when this code is shipped? Do we need to instruct users to do something to upgrade their saved data? If there is no upgrade path possible, how bad will that be for users?

Testing Considerations

Every PR should of course come with tests of its own functionality. What additional tests are still needed beyond those unit tests? How does this affect CI, other test automation, or the testnet?

Compatibility Considerations

Does this change break any prior usage patterns? Does this change allow usage patterns to evolve?

Upgrade Considerations

What aspects of this PR are relevant to upgrading live production systems, and how should they be addressed?

Include *BREAKING*: in the commit message with migration instructions for any breaking change.

Update NEWS.md for user-facing changes.

Delete guidance from pull request description before merge (including this!)

[erights]

This is an orthogonal observable change from #3032 . I claim it is an improvement, in that if the parent is already known to be cancelled, then the child is born cancelled. You could adopt this observable change by itself, without the rest of the PR.

Btw, this observable change did not affect any tests.

[erights]

Turns out that isKnownCancelled's use of this piece of static mutable state does make it a communication channel. This SECURITY BUG seems inherent in this approach, which is why this is a failed experiment. See below.