Summary

Follow-up to #329. On Reyden-backed warehouses, the analytics plugin's stash-full fallback target (EXTERNAL_LINKS + ARROW_STREAM) does not exist — Reyden returns HTTP 501 NOT_IMPLEMENTED: "ExternalLinks disposition is not yet implemented." — so any stash overflow on a Reyden app would surface that 501 to the caller with no further recovery.

This PR adds brief FIFO backpressure inside the stash so transient capacity pressure is absorbed by waiting for the next /arrow-result drain (typically tens of ms — the stash is drain-on-read) rather than immediately escalating to EXTERNAL_LINKS. The existing EXTERNAL_LINKS fallback is preserved for true sustained overload on warehouses that accept it.

What's the fix

• InlineArrowStash: new putWaitMs option and async putBlocking(userId, bytes, signal?). On overflow it waits FIFO up to putWaitMs for a take() / TTL-gc to free a slot, then retries. Honors AbortSignal (aborted request drops out of the queue immediately). When the wait elapses without capacity, returns null and the caller falls back as before.
• AnalyticsPlugin: stash now constructed with putWaitMs: 500; the call site awaits putBlocking and passes the request's abort signal.
• Wait-queue is strictly FIFO. wakeWaiters() walks from the head and shifts each fitting waiter off before invoking it, so a re-entrant gc() inside the waker can't pick up the same waiter twice.

Surface area is small: the synchronous put() is unchanged, the EXTERNAL_LINKS fallback path is unchanged, and putWaitMs = 0 (the default) reduces putBlocking to the synchronous behavior.

Why not the alternatives

• Probe warehouse capability up-front, skip EXTERNAL_LINKS on Reyden: more state, more code, doesn't help if a different warehouse changes behavior later.
• Catch the 501 and surface a clearer error: improves diagnostics, but the request still fails — backpressure prevents the failure.
• Drop the stash entirely and content-negotiate Arrow bytes inline in the HTTP response: structurally cleaner, but a wire-protocol shift better done as a separate PR.

Tests

inline-arrow-stash.test.ts (+8 cases):

• putBlocking succeeds immediately when capacity is available
• waits for a take() to free a slot, then succeeds
• returns null when putWaitMs elapses without a slot freeing
• preserves FIFO order across waiters
• rejects later waiters when the head consumes the freed capacity
• settles null when AbortSignal aborts mid-wait
• pre-aborted signal short-circuits (no wait)
• putWaitMs=0 (default) behaves like the synchronous put()

analytics.test.ts: existing "stash full → EXTERNAL_LINKS" integration test now stubs putBlocking directly instead of put, so the suite doesn't bake the 500ms wait into the run time.

Full suite: 2,682 tests, all green (was 2,674 on the base).

Stack

• feat: decode inline Arrow IPC + warehouse-compat fallback #329 — inline Arrow IPC decode + warehouse-compat fallback (base for this PR)
• (this PR) — backpressure the stash so the fallback isn't reached on Reyden under transient pressure

Targets stack/arrow-3-inline-arrow-fix; will rebase onto main after #329 merges.

Test plan

Verified live against e2-dogfood Test Reyden (warehouse_type=REYDEN, id 0000000002e475d4):

• JSON_ARRAY caller: single SSE result event with row data, no fallback (unchanged)
• ARROW_STREAM caller: single SSE arrow event → GET /arrow-result/inline-<uuid> returns 776 bytes of valid Arrow IPC, application/vnd.apache.arrow.stream (unchanged — new code only activates under capacity pressure)

The new code path (backpressure under contention) is covered by the unit tests; reproducing it live requires saturating the stash, which is outside the scope of this PR.

This pull request was AI-assisted by Isaac.