Compiling the capability survey turned up documentation and comments
that
were true once and now disagree with what the code does. Each was
re-verified against the current source before being changed here;
several
were load-bearing enough to send a reader the wrong way.

Claims corrected:

- docs/hardware/chip-architecture.md called the AIC_CTRL MMIO window
  Device-nGnRnE. It is nGnRE, proven from driver source in
  docs/hardware/mmio-performance.md, and the distinction is what makes
  the
  STR/LDR asymmetry there possible.
- docs/worker-manager.md showed the dispatch wait as sleep_for(50us) and
  called it "not busy-wait". The code spin-polls and cites codestyle
  rule 5 for why it must, sampling child liveness on a steady-clock
  period. Its pseudocode also dropped the check_child_death() call that
  ends the wait with ENDPOINT_FAILURE, so as written the loop would have
  spun forever on a dead child.
- docs/dynamic-linking.md claimed the AICore ELF is registered on every
  launch_aicore_kernel() call; registration is lazy and the handle
  cached.
  Its flow diagram also had the launch order backwards: AICore launches
  before the AICPU Run task, via rtKernelLaunchWithHandleV2, with no
  intervening stream sync.
- docs/comm-domain.md described the window as VMM plus shareable-handle
  import. a2a3 prefers Fabric V2 handles and falls back to VMM IPC; a5
  uses VMM shareable handles only. Peer access via
  aclrtDeviceEnablePeerAccess is still accurate and stays.
- docs/comm-domain.md said a producer CoreCallable declares the SDMA
  workspace requirement. No such API exists; the demo provisions SDMA
  with
  Worker(enable_sdma=True).
- docs/a5-sdma-overlay.md called the a2a3 SDMA path "always on". The
  provider is always compiled; provisioning is opt-in and defaults off.
- The a2a3 SDMA fault-teardown investigation called
  sdma_async_completion_demo unaffected by that change, but the demo
  sets
  enable_sdma=True and so carries the same teardown cost.
- src/common/worker/pto_runtime_c_api.h and the sim device_runner_base.h
  both documented a CallConfig block_dim knob. CallConfig has no such
  field and says so explicitly.
- src/common/aicpu_loader/README.md resolved one symbol rather than
  every
  declared one, put device_id in the per-task KernelArgs when it reaches
  the device once via InitArgs, and gave a per-runtime dispatcher path
  when the SO is shared per arch under build/lib/<arch>/dispatcher/.
- A cpput comment claimed a5 is counter-only; a5 registers SDMA
  event-record and URMA event-handle backends too.
- docs/aicpu-kernel-launch-mechanisms.md credited PR hw-native-sys#537 with
  attempting
  Path B. It is merged and shipped Path A.

Also removes seven references to files under .docs/, which is gitignored
and absent from every checkout, so no reader could follow them. The
surrounding comments now state the invariant that survives on its own
rather than recording where a decision was taken: comments.md keeps
change-rationale and edit history out of code, so a pull-request number
is no more appropriate there than the unresolvable path was. The two
remaining .docs/ mentions are left alone: docs/ci.md lists it as a
NON_CODE path set member, and the review-pr skill creates .docs/review/
as its own scratch directory.

Comments and prose only — no statement or expression changes.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>