D736 — A large KB is linted on every core, read once per request burst, and drawn at a fixed rate

Decision. The whole-KB lint runs its per-concept checks in parallel, one worker per core, and the conformance cache computes each key once however many requests ask for it at the same time (singleflight). The reads an Atlas page repeats on every call are kept off the files: the concept panel's "used by" artifacts are cached on the lint cache's key, the check catalogue reads the types from the graph cache, kb_status reads every concept once and reconciles the search index once per call rather than once per reported miss, and gitx remembers the branch while .git/HEAD holds the same bytes. The lint cache's stamp no longer counts the server's own state under .cartographer/ (search index, read and search-miss telemetry, repair log): every concept_read and every search used to invalidate the whole-KB lint. In the browser the 3D simulation ticks at a fixed 60 Hz, the warm-up that holds the first frame of a new graph stops after 120 ms, and the graph no longer waits for the overview to land.

Why. On a KB of about a thousand concepts the Atlas waited seconds for work it repeated: switching KB fetched the overview, the artifacts, the lint and the status together, and on a cold cache each one ran its own whole-KB lint on a single core (2.3 s) while the graph waited for all of them; the Health panel paid a git process per call and a whole-index reconcile per reported search miss; a 120-144 Hz display ticked the simulation once per frame, doubling its CPU and the pace of a drift tuned at 60 Hz. The cost is concurrency in the lint: every per-concept check must now be free of shared mutable state, which is what surfaced the unguarded markerCache (a crash waiting for two concurrent lints, now behind a mutex).

Alternatives rejected. - Warm the caches in the background after each write: the first reader would still wait whenever a write lands just before it, and the server would lint KBs nobody is looking at. - Run the simulation in a Web Worker: it removes the main-thread cost but doubles the position traffic and reopens the one-writer rule of D234 for a gain the fixed rate already mostly gives. - Tick at 30 Hz once the graph rests: the remaining idle cost is the many-body force, but the drift loses smoothness; proposed in the live session and declined. - Memoise the remote URL on .git/config's stat: git remote get-url also applies insteadOf rewrites from the global config and from included files, which no stat of .git/config sees; the process stays. - Cache git log by HEAD for Activity and Health: --since filters on the committer date while the parsed commits carry the author date, so an in-memory filter would not answer the same question.

Consequences. A new per-concept lint check runs concurrently with the others: it may read the run's shared data but must not write it (the target body cache is the one shared map, behind its own mutex), and go test -race ./internal/lint is how to check it. Findings are produced in no particular order, as before (they came from a map walk). A panic in a check is re-raised on the request's goroutine, so it fails one request, not the process. A file the lint starts reading under .cartographer/ must be added back to the stamp (lintInputsStamp). The branch memo is skipped when .git is a file (worktree, submodule).