[santoshkumarradha]

Two cmd/codeaf tests failed at cleanup with TempDir RemoveAll cleanup: unlinkat …: directory not empty, only under load, never alone. Neither test starts a goroutine of its own — the first one's
whole body is a t.Setenv("CODEAF_HOME", t.TempDir()), a cache write and a read-only gate — so the
writer was never the test's.

The writer, named

warmV3Models is started fire-and-forget by guard.Go, which joins nothing at shutdown, and it
finishes by writing the picker's cache through tui3.WriteModelCache — which resolves CODEAF_HOME
at the moment it writes, not when the warm started. On a cold cache the warm is a network
round-trip, so a warm can outlive the conversation that asked for it and land its write in whichever
state root is current when the rows finally arrive: in a package run, the next test's t.TempDir(),
whose clean-up then fails with directory not empty. That is both symptoms at once — only under load
(slow enough to cross a test boundary) and never alone (no earlier test to leak one).

v3Process.closeAll's own header promises that nothing this process started is still writing under
its profile once it closes. For this writer the promise was not true.

The fix, in two parts

Seat the warm on the tracker it should always have been on. poolErrandGoCtx is poolErrandGo
for an errand that must observe cancellation rather than only be waited for; v3Process.warmModels
starts the warm through it, so closeAll cancels the wait and joins the goroutine, exactly as #1124
did for the pool errands. Both call sites move: the chat door (chatv3.go:419) and the engine
(engine.go:708).

Give the warm reads that cannot block. The wait now belongs to Catalog.Warmed(ctx), the one
context-observing door the close can end; after it answers true the rows are in hand, so the two
reads behind it — the context window and FetchedAt — must not ask blocking questions of a catalog
the errand has just stopped waiting on. ContextLengthNow and FetchedAtNow read rowsNow (the
ModelsNow door) and answer zero / the zero time while a lazy catalog is still warming.

ContextLengthNow is byte-for-byte the id matching of its blocking twin — byID[normalizeID(id)],
the same normalisation of the reasoning-effort suffix and a leading ~. That matters: the obvious
cheaper route, reusing the launch's existing non-blocking v3ContextWindow seam, matches with
EqualFold, so a conversation started on model:high would answer a zero window — the exact
number the warm exists to learn. The cell caught that in its own first draft and said so.

Consequence, stated rather than buried

A process that closes before its catalog resolves now writes no picker cache for the next launch,
where before the orphaned warm eventually wrote one. That is the intended trade: the old write was
the bug — it landed in a profile the process no longer owned. The warm runs again at the next launch.

Proof the test bites

A test for a race can pass by asserting nothing, so I checked it by mutation rather than by reading.
In a clone of this branch, unseating the warm back to its old shape —

go func() { warmV3Models(context.Background(), models, agent, started) }()

--- the new test with the seating REMOVED (expect FAIL) ---
        profile's errand tracker lands in whichever root is current when the rows arrive — the
        TempDir clean-up's own "directory not empty". It must be seated on the tracker and joined
        at close (chatv3_process.go, poolindex.go).
FAIL	github.com/Agent-Field/codeaf/cmd/codeaf	0.211s

--- reverted, expect ok ---
ok  	github.com/Agent-Field/codeaf/cmd/codeaf	6.560s   (-count=3)

The test makes the race certain rather than likely: its endpoint holds the catalog fetch open until
the test releases it, so an unseated warm is certainly still waiting at the close, and the test has
already moved CODEAF_HOME to a directory it owns before releasing it. Deterministic, at -count=3.

Evidence

From a fresh clone of this head (c438a8f75b079f7f068d97a55dd289660ee23792) on the Spark
cluster — log ~/src/trees/check-c438a8f75b079f7f068d97a55dd289660ee23792.log, HEAD on its first line:

check	result
gofmt -l ./internal ./cmd	nothing printed
go vet ./internal/... ./cmd/...	clean
go test ./internal/guard ./internal/namelaw	ok
go run ./cmd/codeaf-changes check	82 entries, all well formed
go test ./cmd/codeaf -count=1	ok 123.8s
go test ./internal/catalog -count=1	ok 0.15s

The catalog package is in the list because this adds two doors to it.

🤖 Generated with Claude Code