[justinchuby]

What this is

The first bounded step of docs/architecture/SESSION_CONCURRENCY.md §13
Phase 1's state split: WorkflowRuntime's flat field list becomes the owners
§3 already names. W = 1 structural prerequisite only — this PR makes
nothing concurrent and claims nothing about concurrency.

Before, the runtime held the compiled plan, the ORT and native backend
handles, the caches derived from them, every counter and the interpreter's
session cells in one list, so which §3 category a field belonged to was a
fact about the comment above it. A future worker could be handed state it
must not share without anything failing to compile.

The split

Owner	Category	Holds
WorkflowPlan behind Arc	§3.1 immutable	workflow, compiled graph, derived output sets, memory plan, backend, adapter service, preprocessing, speculative contract
WorkerBackend	§3.2 backend handles	component sessions, execution islands, native component set
WorkerRuntimeState	§3.2 storage, §3.3 access	bindings, allocators, device values, embedding tables, session cells, session leases, iteration variants, adapter cache, counters
WorkflowPass	§3.3 per-execution	the pass id and its telemetry, created when a pass opens and dropped when it ends

WorkflowPlan carries a static assertion that it is Send + Sync — a
mutable cell added to it fails the build rather than quietly making the plan
unshareable. Both worker types are structurally !Send/!Sync through a
PhantomData marker: no new unsafe impl, no blanket lock, and static
assertions keep it that way. Engine's existing unsafe impl Send is
untouched.

Id namespaces

Made explicit, because a future worker minting from the wrong one hands two
conversations the same id.

• Worker-local: PassId (was the bare workflow_execution_generation
  counter) and GraphCaptureId. Both are only ever compared against state
  the same worker owns. Component capture ids came from bindings.len(),
  which is a counter only for as long as nothing is removed from that map;
  islands seeded a hand-rolled island_id * 1000 stride, kept because it
  makes a captured graph's id say which island captured it. Exhaustion is now
  a typed error rather than a wrap into ORT's -1 uncaptured sentinel.
• Shared: session ids are the only ones that leave the engine, so they
  mint from engine::ids::SharedSessionIds, an atomic allocator. Ids still
  start at 1 and still increment by 1 — nothing externally observed
  changes.

#2019's contracts are preserved

WorkflowRuntime::drop still clears island bindings, then values, then
allocators: a Value is a bare OrtValue handle with no back-reference to
its allocator, so nothing in the type system orders that release. The clears
moved into WorkerRuntimeState::release_ort_state, called from the same
Drop. worker is now declared before backend, and _ort_environment
stays last, so the field order agrees with the explicit teardown instead of
contradicting it.

Explicitly not in this PR

No N > 1. No pre-enqueue lease. No server routing. EngineOwner's
unsafe impl Send is not removed. Batching semantics unchanged. No public
API change — the split is private to the crate and every accessor keeps its
signature.

Tests

New, in pipeline/runtime_state.rs, pipeline/mod.rs and engine/ids.rs:

• static: the plan is Send + Sync behind Arc; WorkflowRuntime is
  neither Send nor Sync; each worker type and id allocator likewise.
• the plan really crosses a thread boundary behind an Arc and reads the
  same on the other side.
• pass ids start at "none", are fresh per pass, and an iteration variant
  mints in its own namespace without advancing its parent's.
• capture ids are unique and non-negative for component bindings, disjoint
  across islands, and exhaustion is refused rather than wrapped.
• session ids start at 1, never repeat, and are collision-free across eight
  concurrent minters.
• a conversation still continues from — and is still released through — the
  worker's own cells, and the inner session lease still refuses a second
  holder by name.
• the drop path still releases cached values, before the allocators that own
  their memory.

Verification

cargo fmt --all -- --check                                              # clean
cargo clippy -p onnx-genai-engine --all-targets -- -D warnings          # clean
cargo clippy -p onnx-genai-engine --all-targets --features native-backend -- -D warnings  # clean
cargo clippy -p onnx-genai-server --all-targets -- -D warnings          # clean
cargo test -p onnx-genai-engine --lib                                   # 494 passed, 1 ignored
cargo test -p onnx-genai-engine --lib --features native-backend         # 637 passed, 1 ignored
cargo test -p onnx-genai-server --lib                                   # 273 passed, 2 ignored
cargo test -p onnx-genai-engine --test workflow_session_continuation    # 16 passed
cargo test -p onnx-genai-engine --test multi_session                    # 2 passed
cargo test -p onnx-genai-engine --test static_cache_workflow            # 7 passed
cargo test -p onnx-genai-engine --test workflow_performance_conformance # 1 passed, 2 ignored
cargo test -p onnx-genai-engine --test authored_body_selects_executor \
  --test authored_iteration_executors --test model_package \
  --test batching_capability                                            # all passed
python3 scripts/check_feature_gate_coverage.py                          # pass
python3 .github/scripts/verify_cuda_test_honesty.py --source-scan       # pass

cargo clippy --workspace cannot run in this worktree: git worktree add
does not populate the crates/onnx-runtime-cpuinfo/vendor/cpuinfo
submodule, which is an environment fault rather than a result. The crates
this change touches are covered above.

Docs

SESSION_CONCURRENCY.md §13 Phase 1 gains a note that the state split
proper has landed for the interpreter, and that the phase is not done:
the Box<Session> holders and the thread_affinity extension to the native
handles remain. Citations into pipeline/mod.rs that the move invalidated
are repointed at pipeline/runtime_state.rs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 80.49%. Comparing base (398cff8) to head (a2bebe6).
⚠️ Report is 13 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2044      +/-   ##
==========================================
+ Coverage   80.25%   80.49%   +0.23%     
==========================================
  Files         409      425      +16     
  Lines      189791   209135   +19344     
  Branches   189791   209135   +19344     
==========================================
+ Hits       152319   168334   +16015     
- Misses      32048    35141    +3093     
- Partials     5424     5660     +236     

Flag	Coverage Δ
cli-ort-linux	72.51% <ø> (?)
cli-ort-windows	72.10% <ø> (?)
mlas	85.10% <ø> (?)
offline	80.62% <ø> (+0.37%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 67 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.