yarn typecheck reports phantom and stale errors because of the incremental tsbuildinfo cache

[pat-lewczuk]

Hit by two independent agent runs adding route-handler tests. It produces false results in both
directions, which makes it worse than a plain slow build.

What happens

yarn typecheck is tsc --noEmit with incremental caching into tsconfig.tsbuildinfo (untracked,
at the repo root). Across branch switches and file moves/renames the cache goes stale, and then:

• it reports phantom errors on a clean tree — observed as TS2451 duplicate-identifier errors for
  files that no longer exist at the paths the cache remembers;
• and it keeps reporting an error after the fix has landed, so a correct tree looks broken.

Both were observed in the same session, on different branches.

Why it matters more than a normal caching wrinkle

yarn typecheck is one of five validation.commands in .ai/agentic.config.json — it is the gate
every change is measured against, and the one that substitutes for yarn lint (which cannot run at
all, see #179). A gate that intermittently fabricates failures trains everyone, human and agent, to
distrust it or to work around it with reruns.

Reproduction is branch-history dependent rather than a fixed recipe: switch between two branches whose
test files differ in name or location, then run yarn typecheck. The workaround is:

rm -f tsconfig.tsbuildinfo && yarn typecheck

Options

1. Clean in the script — "typecheck": "rm -f tsconfig.tsbuildinfo && tsc --noEmit". Costs a few
   seconds per run, removes the whole class of problem. Probably right for a gate command.
2. Drop incremental: true from tsconfig.json — same effect, but also affects editors and any
   other tsc consumer.
3. Keep incremental, move the buildinfo per-branch — more machinery than this is worth.

Option 1 is the smallest change that makes the gate trustworthy. Also worth confirming
tsconfig.tsbuildinfo is gitignored (it is currently untracked but should be explicit), the same
hygiene question as #131's next-env.d.ts.