fix: bound memory during device output streaming (Diagnostics/backup logs)

[marlonbarreto-git]

Summary

• Diagnostics/backup log streaming floods the main queue → unbounded memory growth #36
  Fixes Diagnostics/backup log streaming floods the main queue → unbounded memory growth #36. Live syslog/backup streaming enqueued one DispatchQueue.main.async block per pipe chunk from a background readabilityHandler. When a device produced output faster than the main thread consumed it (e.g. verbose iPhone syslog), the main-queue backlog of un-executed blocks grew without bound — each retaining a String — exhausting memory (observed tens of GB before the machine froze).

What changed

A single-in-flight, coalesced main-queue flush backed by two small, bounded, UTF-8-safe buffers:

• CoalescingOutputBuffer (byte-level) — used in Shell.runStreaming and PyMobileDevice.runStreaming. Preserves raw text, including the \r progress updates emitted by idevicebackup2/pymobiledevice3, so it is safe for every caller (not just \n-based syslog). Bounds buffered bytes and holds back a multi-byte scalar split across two reads (previously String(data:encoding:.utf8) returned nil and dropped the whole chunk).
• StreamingLineBuffer (line-level) — used in DiagnosticsManager.startSyslog to split coalesced text into whole lines with a bounded ring buffer. Fixes multi-line chunks that counted as a single entry and the 5000-"line" cap that actually bounded 5000 chunks.

Also: retain the fallback idevicesyslog Process so stopSyslog() can terminate it, and clear readability handlers on launch failure.

The key property: the number of queued main-queue blocks is bounded to one in flight, regardless of producer rate.

Flow (before → after)

sequenceDiagram
    participant Dev as Device (fast producer)
    participant BG as readabilityHandler (bg)
    participant Main as Main queue (slow consumer)
    Note over Dev,Main: BEFORE — one main.async per chunk
    Dev->>BG: chunk 1..N (thousands/sec)
    BG->>Main: main.async { onOutput } × N (unbounded backlog)
    Note over Main: backlog grows → memory exhausted
    Note over Dev,Main: AFTER — coalesced single-in-flight flush
    Dev->>BG: chunk 1..N
    BG->>BG: append to bounded buffer
    BG-->>Main: schedule ONE flush (if none pending)
    Main->>Main: drain buffer → deliver (backlog ≤ 1)

Loading

Verification

• Flood test against the real patched Shell.runStreaming: 3,000,001 lines with a deliberately slow consumer peaks at 13 MB RSS (previously unbounded). All lines observed (no loss).
• Unit tests for both buffers (line splitting, bounded capacity keeps newest, UTF-8 scalar split across reads, \r preservation, drain/finish) — added under Tests/PhosphorTests/.
• Scripts/build.sh: all 32 regression checks pass; release build succeeds.

Test Plan

• Connect a verbose device, open Diagnostics → Logs, confirm memory plateaus over several minutes.
• Confirm backup progress (idevicebackup2/pymobiledevice3) still updates.
• swift test (unit tests) and Scripts/build.sh (regression) green.

[marlonbarreto-git]

Hi @momenbasel @AJV20 — thanks for building Phosphor. I opened a few issues (#36–#40) and a draft PR (#41) fixing #36, the unbounded-memory issue in Diagnostics/backup log streaming. #41 follows CONTRIBUTING (regression checks + build pass) and includes a flood-test showing peak RSS drop from unbounded to ~13 MB. Since #33 already reworked the syslog fallback, your eyes there would be especially welcome. Happy to adjust anything. Thanks!