fix: hold GET memory reservation for the whole streaming-response lifetime

[ServerSideHannes]

What

Fix the dominant s3proxy concurrent-backup OOM: streaming GET responses ran completely ungoverned.

Root cause (reproduced + profiled locally)

A multipart-encrypted GET returns a StreamingResponse whose body is sent after the request handler returns. The handler released the GET's memory reservation in its finally — i.e. before a single byte was streamed — so concurrent streaming GETs ran with no governance. Each stream holds ~one decrypted 8MB frame in uvicorn's transport send buffer, so a flood of concurrent GETs accumulated N × 8MB and OOMKilled the pod (exit 137) while the limiter's active_mb read ~64MB.

This is what kept 2026.6.13 OOMing in prod: the kill windows were dominated by 100+ concurrent GETs, and the earlier copy-govern fix (#97) never touched this path.

Profiler evidence (S3PROXY_TRACEMALLOC): peak snapshot held crypto.py:394 (aesgcm.decrypt) = 576–720MB across 72–90 live 8MB frames = 90% of tracked memory.

Fix

• Transfer the reservation to the streaming body iterator (_release_after_stream) so it's released only when the stream is exhausted or the consumer disconnects. The limiter now bounds how many streaming GETs run at once (admission control).
• Drop the per-frame nested try_acquire_memory in the multipart GET path: concurrency is now bounded at admission, so the working set is O(concurrent streams) not O(frames), and a nested per-frame acquire would deadlock against the held reservation.

Proof (local, prod config: 512Mi cap / 64MB budget)

Test	Before	After
90-concurrent multipart GET flood	OOMKilled, exit 137, 0/180	180/180 ok, peak ~325MiB
realistic upload+GET mix	—	106/106 ok, peak ~305MiB

Profiler with fix: tracked memory 812MB → 112MB, live frames 90 → 11.

Scope / follow-up

A synthetic 192-parallel-UploadPart torture test (far harsher than prod's ~57 PUTs) can still OOM via untracked uvicorn C-level socket receive buffers for concurrent PUT bodies — a separate, lesser concern (candidate: uvicorn --limit-concurrency). Not addressed here.

Tests

• New tests/unit/test_streaming_get_reservation.py: pins that the reservation is held for the whole stream and released exactly once (incl. early consumer disconnect).
• Full unit suite green (451); ruff check + ruff format --check clean.