M6: Version bump rehearsal, LLVM 24

[tamnd]

Four weeks, no new lessons. LLVM ships twice a year and the 24.x branch is expected in late January 2027, so a major version bump lands mid project no matter when the project starts. The point of scheduling it rather than reacting to it is to measure what a bump costs while the corpus is still small enough to fix.

Exit criterion. A written cost report giving the measured cost of a major version bump in hours, split by category: citations that moved, generated sections that changed, claims that became false, lessons that need rewriting, and toolkit parsing that broke.

Tasks

• driftbot's release note triage document for LLVM 24
• Every citation revalidated against the new tag
• Every generated blueprint section regenerated and diffed
• Every notebook re-executed against the 24 toolchain
• The cost report

Gates

• The cost report exists with real numbers in it
• Both the 23 and 24 versions of the site are live under mike and the reader can choose

If the cost exceeds three weeks of work, then the citation density or the generation ratio is wrong and M6 ends with a structural change rather than a shrug. That is the whole reason for doing this early.

The pin does not necessarily move to 24 here. That decision gets made with the cost report in hand. What is not negotiable is doing the exercise.