host_build_graph: route boot failure through common teardown; fix Level-1 docs

[ChaoZheng109]

Summary

Follow-up on unaddressed review feedback from #1452. Fixes the three findings tracked in #1515 — one correctness bug and two stale-doc drifts.

1. (Bug, Major) Boot-thread failure bypasses AICore teardown and the finished-barrier

aicpu_executor.cpp — both boot-failure branches in run() returned -1 directly. Since #1452 made every AICPU thread schedule, the boot thread (thread_idx == aicpu_thread_num_ - 1) now owns a non-empty core slice (handshake_partition assigns it [lo, total) in init()), so an early return skips the common teardown:

• shutdown(thread_idx) — the boot thread's AICore cores never get their exit signal and spin forever on an unclosed register window.
• finished_count_ — tops out at aicpu_thread_num_ - 1, so finished_ never publishes and runtime_destroy never runs; the host hangs into the op-execute timeout (507018), masking the real error.

Fix: record the failure in run_rc, leave rt null (the dispatch block already skips on rt == nullptr), publish runtime_init_ready_ once so peers stop spinning, and fall through to the common teardown. Success-path behavior and the task-count-latch-before-release ordering are unchanged.

2 & 3. (Docs) profiling_levels.md Level-1 describes an on-device orchestrator

host_build_graph boots scheduler-only (orchestrator runs on the host), so the device log has no orch_* lines and no "PTO2 total submitted tasks" line. The Level-1 section still listed them, used the count formula N_sched*2 + N_orch*1 + 1, and showed a device-orch example capture; the summary table listed Level-1 = 7.

Fix: rewrite the section scheduler-only (N_sched*2, N_sched == aicpu_thread_num) with a 4-thread example, and correct the summary-table count 7 → 8.

Testing

• Incremental build of the a2a3 onboard host_build_graph AICPU runtime — aicpu_executor.cpp compiles and libaicpu_kernel.so links cleanly.
• markdownlint-cli2 passes on profiling_levels.md.
• The Multi-threaded AICPU Scheduler with Parallel Task Dispatch #1 fix is on the boot-failure (error-injection) path — a natural onboard repro is impractical; verified by inspection against the common-cleanup contract in run().

Fixes #1515.

🤖 Generated with Claude Code

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 8c32a8f8-c68d-4f68-832e-16025e8e58e5

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The runtime boot path now handles attachment failures through shared teardown and publishes initialization readiness in all boot outcomes. Profiling documentation now describes scheduler-only Level 1 execution and updates its log examples and counts.

Changes

Host build graph boot and profiling

Layer / File(s)	Summary
Boot failure teardown src/a2a3/runtime/host_build_graph/aicpu/aicpu_executor.cpp	Boot failures set failure state without returning early, publish runtime_init_ready_, skip runtime dispatch, and continue through scheduler shutdown and completion handling.
Scheduler-only profiling documentation src/a2a3/runtime/host_build_graph/docs/profiling_levels.md	Level 1 documentation, examples, and LOG_INFO_V9 counts now contain scheduler-only timing and summary output.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Sequence Diagram(s)

sequenceDiagram
  participant BootThread
  participant RuntimeAttachment
  participant WaitingThreads
  participant SchedulerContext
  BootThread->>RuntimeAttachment: Check prebuilt_arena and attach_populated
  RuntimeAttachment-->>BootThread: Runtime or boot failure
  BootThread->>WaitingThreads: Publish runtime_init_ready_
  BootThread->>SchedulerContext: shutdown(thread_idx)
  BootThread->>BootThread: Increment finished_count_

Loading

Possibly related PRs

• hw-native-sys/simpler#1452: Updated the same runtime boot and orchestration control flow.

Poem

A bunny found a boot path bright,
No early exits fled the night.
Schedulers count, the logs now sing,
Readiness wakes each waiting thing.
Teardown hops in tidy flight!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly summarizes the two main changes: boot-failure teardown routing and Level-1 documentation updates.
Description check	✅ Passed	The description is directly about the runtime fix and profiling docs, matching the change set.
Linked Issues check	✅ Passed	The code and docs updates match issue #1515's boot-failure teardown and Level-1 profiling requirements.
Out of Scope Changes check	✅ Passed	No unrelated code changes are indicated beyond the runtime fix and the documented Level-1 updates.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}