refactor(backends): route intrinsic generation through adapter_scope (Epic #929 Phase 2)

[planetf1]

Parent epic: #929
Phase: 2
Depends on: #1141
Blocks: #1466 (generate / parse spans only)

Problem

Adapter-function generation does not run inside adapter_scope. LocalFileBinding
can now prepare / activate / deactivate / release a PEFT adapter, and
adapter_scope() demonstrably flips the real model's active adapter set — but the
model call in _generate_from_intrinsic still happens outside that scope, and the
standard generation path deactivates adapters first
(_generate_with_adapter_lock). So the lifecycle verbs and the generation they
exist to serve are not yet connected.

This was #1141's acceptance criterion 66 — "E2E adapter call (prepare →
activate → generate → deactivate → release) passes" — and it is not excluded by
#1141's out-of-scope list. It was deferred during implementation of PR #1454, and
the deferral is disclosed only in a test docstring
(test/backends/test_adapters/test_local_file_e2e.py) which refers to a follow-up
issue that was never created. This issue is that follow-up.

It is also not what the epic informally calls "4.1". 4.1 is #1144 (shim
removal plus the intrinsics_and_adapters.md rewrite), whose scope does not
mention generation, activation, scoping or locking. Without a separate issue this
work falls through the gap between the two.

Known constraint: the activation lock is not reentrant

_adapter_activation_lock returns _generation_lock, which is not reentrant.
Once generation runs inside adapter_scope and activate() re-acquires that
lock, it deadlocks. This was raised in PR #1454 review. Separately, the lock is
weaker than its docstring implies today — it does not serialise everything the
docstring claims — so the fix should reconcile behaviour and documentation rather
than only unblocking the deadlock.

Any solution needs to keep the current guarantee that a caller cannot observe
another caller's adapter active, which is what the existing
deactivate-before-generate behaviour buys.

Scope

• Route the model call in _generate_from_intrinsic through adapter_scope, so
  the prepared and activated adapter is the one used for generation.
• Resolve the activation/generation lock reentrancy, and align the docstring with
  actual behaviour.
• Extend the existing e2e coverage so the full lifecycle is exercised with
  generation inside the scope, and remove the deferral note from
  test_local_file_e2e.py.

Out of scope

• EmbeddedBinding (feat(backends): EmbeddedBinding implements apply_activation (Granite Switch path); remove render_controls + set_request_adapter (Epic #929 Phase 2) #1142) and ServerMediatedBinding.
• Span emission for generate / parse — that is blocked on this issue and
  tracked in feat(telemetry): add complete adapter-function lifecycle observability #1466.

Acceptance criteria

• _generate_from_intrinsic performs its model call inside adapter_scope
• Generation demonstrably uses the activated adapter, asserted rather than
  smoke-tested
• No deadlock: activate() inside adapter_scope with generation nested
  completes, with a regression test that would deadlock under the current lock
• _adapter_activation_lock / _generation_lock docstrings match observed
  behaviour
• A concurrent-callers test shows one caller cannot generate against another
  caller's adapter
• test_local_file_e2e.py covers prepare → activate → generate → deactivate →
  release, and its deferral note is gone
• uv run pytest test/ -m "not qualitative" passes; the e2e test passes on
  GPU hardware

References

• feat(backends): LocalFileBinding implements verbs (PEFT/aLoRA path) + from_catalog() (Epic #929 Phase 2) #1141 acceptance criterion 66 — the criterion this issue carries
• PR feat(backends): LocalFileBinding implements verbs (PEFT/aLoRA path) + from_catalog() (Epic #929 Phase 2) #1454 — where the deferral happened, and the review comment identifying the
  deadlock
• refactor(backends): AdapterMixin verb rename/narrow + resolve_model_options + IntrinsicMetricsPlugin (Epic #929 Phase 2) #1140 — introduced adapter_scope and the verb set
• refactor(intrinsics): remove deprecation shims; rewrite intrinsics_and_adapters.md; write 3 tutorials (Epic #929 Phase 4) #1144 — "4.1"; explicitly not this work

[planetf1]

Heads-up from PR #1454's verification pass: #1528 is worth reading before this issue picks a lifetime for its bindings.

LocalFileBinding.release() unloads the weights but does not deregister — unload_peft_adapter clears the backend's loaded set, while _added_adapters (the registered set that add_adapter checks for duplicates) keeps its entry, and there is no inverse verb to call. So a released qualified_name stays claimed for the backend's lifetime and no later binding can register under it.

That bites here specifically if this cutover ends up preparing a binding per call rather than once per session — the second call cannot register. #1528 frames the prior question: the merged WeightsBinding ABC says release() is terminal and the binding is not reusable, while #1141's text promised a release()+prepare() cycle. Worth settling which is intended before choosing a binding lifetime here, even if the answer is "terminal is correct, prepare once".

[planetf1]

Heads up: #1528 is deciding whether a WeightsBinding is reusable after release(). Worth a quick check before this issue fixes a binding lifetime, per #1528's own note ("worth resolving before #1465 picks a lifetime for its bindings").