perf(run): the up-to-date check still walks the source tree on every run (~0.85s in a large consumer)

[MarkMichaelis]

Follow-up to #519 / #546, recorded at session wrap-up so it does not live only in a conversation.

What is left

With the launch plan cached (#546), an unchanged run of a sample app costs ~0.26s, of which the application itself is ~0.10s. The largest remaining launcher cost is the up-to-date check: Test-BuildRequired walks the pruned source tree on every run to find the newest source timestamp.

Measured in a large consumer (MarkMichaelis/CodiwomplerSocialMedia, ~6k source files after pruning): ~0.85s per run, plus ~0.7s for project discovery (Find-RunnableProjects, which reads every .csproj). Those two are now most of the launcher's wall clock there.

Why it is not already solved

The walk is what makes --no-build unnecessary in the common case, so it cannot simply be dropped. #521 added a trailing --no-build that skips it when the developer knows the build is current, which is a manual escape rather than a fix.

Possible directions (none chosen)

• Cache the discovered runnable project the way the launch plan is cached, invalidated on solution/project-file changes -- removes the .csproj-reading pass.
• Short-circuit the timestamp walk: stop at the first source newer than the build output instead of finding the newest (needs the threshold pushed into the walk).
• Compare against a directory-level watermark, or use the build output's own up-to-date check.

Verified vs assumed

Verified: the timings above, measured directly. Assumed: that the directions listed are feasible -- none has been prototyped.

[MarkMichaelis]

Priority: P3 (priority/important-longterm) -- set by a session at filing, not confirmed by the owner.

Reason: the owner has twice asked for launcher start-up to be faster (#519, #546), so this is squarely in an area they care about, and the measurements are already in hand. It is not P2 because the launcher is now fast enough to use comfortably (~0.26s on a small project) and the remaining cost only bites in a large repository, where the trailing --no-build from #521 already offers an escape.

Filed at session wrap-up; nothing started.