feat(insight): add async export scheduling

[wangyb-A]

Summary

• move Workflow Insight rendering, truncation, export, and flush off the checkpoint thread
• coalesce cumulative snapshots with one lazy daemon lane per exporter
• drain and flush with one shared configurable timeout
• prevent one exporter object from being driven by multiple live scheduler lanes

Design

• one lazy daemon worker per exporter object
• one in-flight record and one latest pending record per lane
• a newer pending snapshot replaces the older one, regardless of execution ARN
• scheduling only swaps the pending reference, starts or wakes the worker, and returns
• custom-renderable and non-JSON values are not inspected on the checkpoint thread
• exporter-specific copying, rendering, truncation, and export stay on the worker
• export_timeout_seconds defaults to 5 seconds and bounds caller wait for drain plus flush
• timed-out lanes retain at most one detached flush so buffered records can publish before idle exit
• workers stop when idle; blocked workers are retained and never replaced
• worker-start failures disable the lane without escaping hooks or bypassing invocation cleanup
• exporter ownership lasts as long as its lane, including an orphaned live worker
• no core SDK changes

Validation

• full Insight suite: 133 passed, including local-runner e2e
• scheduler suite: 31 passed
• config and ownership suite: 44 passed
• focused scheduler and async-plugin suite: 40 passed
• latest-pending replacement, 100-record burst, opaque/deep-value scheduling, exporter isolation, shared-timeout, queued/popped detached-flush races, sizing-resource failure, startup-failure, cleanup, and ownership-finalization regressions passed
• the queued-barrier, sizing-resource, deep-record, and concurrent-finalization regressions were proven to fail on pre-fix production code and pass with their fixes
• mypy passed
• Ruff lint and format passed
• wheel and sdist built
• full repository collection completed without errors
• prior Workflow Insight conformance remains CloudWatch 18/18 and S3 18/18
• exact-SHA local correctness review at afc10fe: no remaining or new Critical/High findings

Latest review fixes

• replaced per-execution admission and retained-graph estimation with one structurally bounded latest-pending slot per exporter lane
• removed checkpoint-thread object sizing and custom-state traversal
• retained worker-side support for custom-renderable and non-JSON values
• let resource failures escape JSON sizing into the worker isolation boundary instead of retrying through export
• passed the original record to custom rendering when lane-local container copying fails
• preserved bounded invocation-end wait and one eventual detached flush after timeout
• coalesced canceled and detached flush markers at their rightmost queue position so buffered records remain covered
• preserved later invocations' detached flushes when an older barrier is already in flight or still queued
• retained strong ownership references through lock release so custom exporter finalizers cannot re-enter the ownership lock
• contained worker-start failures so instrumentation cannot break checkpoint hooks or state cleanup
• enforced exporter ownership across plugin instances by lane identity and lifetime

no linked issue: #687 tracks broader Workflow Insight follow-up work

[wangyb-A]

/ai review

[wangyb-A]

/ai review

[github-actions]

Codex AI review · Finding arf_v1_hhdiypndoo4kuhjkuxlfcqckuf

[P2] Release superseded records after dropping the lane lock. Replacing _pending decrements the previous record while _cond is held. Supported transformed/non-JSON values may define __del__, so the checkpoint thread can run arbitrary blocking code here; a finalizer that re-enters this lane deadlocks on the non-reentrant lock. Retain the displaced record until after the with block, similarly defer pending cleanup in _disable_locked, and add a finalizer re-entry regression.

[github-actions]

Codex AI review · Finding arf_v1_54if3osmob4i3ya37qb2nfe6ew

[P2] Keep deeply nested records isolated between exporter lanes. This recursive deepcopy raises for sufficiently deep built-in containers, after which _export_one passes the original canonical record to every lane. Concurrent exporters can then mutate or normalize the same object, producing timing-dependent or corrupted output. Clone built-in containers iteratively while preserving aliases and cycles, and add a two-lane deep-record test where one exporter mutates its input.

[github-actions]

Codex AI review

Two P2 scheduler issues remain in _export_scheduler.py (lines 122 and 53): checkpoint-thread lock safety and cross-lane record isolation.

Reviewed commit afc10fe129032aa3f2496bcbea3af41b17fe11bd. Workflow run

[wangyb-A]

Superseded by #719, which re-authors the async exporter around one plugin worker and one latest-pending slot with a much smaller diff.