Fix/postchannel autoflush permanent stall

[hectorhdzg]

Problem

Under sustained intermittent send failures (e.g. a load balancer recycling
connections and returning occasional 503s), PostChannel could permanently stop
draining its in-memory queue. Once wedged, the queue saturates and telemetry is
silently dropped as QueueFull, and only a process restart recovers it.

Root cause

Auto flush is fire-and-forget, but it routed through flush(), whose async path
defers the send via a 0ms timer and then calls _waitForIdleManager(). That
re-arms _flushCallbackTimer and polls (every FlushCheckTimer) until the
HttpManager reports it is "completely idle" (no outstanding requests AND an empty
send queue) before clearing the timer.

While _flushCallbackTimer is non-null, BOTH the normal scheduled timer
(_scheduleTimer) and any further auto flush are suppressed. Under continuous
intermittent failures there is almost always a retry in-flight or a re-queued
batch, so the manager is rarely completely idle at the polling instant: the wait
never completes, _flushCallbackTimer stays set forever, and the channel stops
draining permanently.

Fix

Keep the original deferred fire-and-forget timing (so queue/discard semantics are
unchanged) but do not park the scheduler waiting for the manager to become idle:

• _flushImpl() gains a waitForIdle parameter. When false (fire-and-forget auto
  flush), the completion logic runs immediately instead of via
  _waitForIdleManager(), so after the send the normal schedule resumes. Explicit,
  user-initiated flush() keeps its existing wait-for-idle behaviour.
• _performAutoFlush() mirrors flush()'s async path (clear scheduled timer, queue
  the batches, send on a 0ms timer) but calls _flushImpl(..., waitForIdle=false).

A busy / non-idle HttpManager can therefore no longer wedge draining, while the
event ordering and QueueFull discard behaviour under overload are preserved.

Tests

Adds a regression test ("auto flush does not wedge behind the flush()
wait-for-idle timer") that fails on the old code (the residual wait-for-idle poll
timer suppresses the next auto flush) and passes with the fix. All existing
queueSize/discard/reset tests remain green (behaviour unchanged).

[Copilot]

Pull request overview

Fixes a PostChannel auto-flush edge case where intermittent/continuous send failures could cause _flushCallbackTimer to remain set indefinitely (due to wait-for-idle polling), suppressing both scheduled sending and subsequent auto flushes and eventually leading to silent QueueFull drops.

Changes:

• Adds a waitForIdle parameter to _flushImpl() so fire-and-forget auto flushes can complete without parking the scheduler behind _waitForIdleManager().
• Reworks _performAutoFlush() to mirror the async flush scheduling behavior while calling _flushImpl(..., waitForIdle=false).
• Adds a regression test to ensure auto flush continues draining across multiple bursts without wedging behind the wait-for-idle timer.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 3 comments.

File	Description
package.json	Updates the tar override minimum version.
common/config/rush/pnpm-config.json	Updates the Rush/pnpm tar globalOverride minimum version.
channels/1ds-post-js/src/PostChannel.ts	Implements the non-wedging auto-flush behavior by avoiding wait-for-idle for fire-and-forget flushes.
channels/1ds-post-js/test/Unit/src/PostChannelTest.ts	Adds a regression test covering the permanent-stall scenario.

[rads-1996]

Is this a genuine concern?

[hectorhdzg]

Not really, tar is only used internally not even a direct dependency in any code in this project, it get dragged with npm

[JacksonWeber]

LGTM