PR-N2: remove DeterministicEngine + DeterministicTokenizer from scheduler tests

[FluffyAIcode]

Why

Second installment of the no-test-doubles cleanup (per the user's principle: fake = mock, all banned). PR-N1 cleared verifier protocol mirrors. PR-N2 clears the scheduler-conftest engine + tokenizer mirrors. PR-N3 will tackle the HTTP shim's separate copies and engine subtypes; PR-N4 will clean up the SDK conftest stub + final CI consolidation.

Scope

In scope (this PR)	Out of scope
tests/inference_engine/scheduler/conftest.py — DeterministicEngine, DeterministicTokenizer classes + their fixtures	tests/inference_engine/server/conftest.py's separate copy (PR-N3)
tests/inference_engine/scheduler/test_scheduler.py — 20 tests using those doubles	HTTP shim tests + engine subtypes (PR-N3)
	test_pooled_verifier.py — exempt; PR-D2 retires PooledVerifier module

What was deleted

File	Δ	What
tests/inference_engine/scheduler/conftest.py	−197/+62 net	DeterministicEngine + DeterministicTokenizer classes (~120 lines) + their fixtures (tokenizer, short/long/slow_engine, reject/queue_scheduler). Slab fixtures (slab_config, small_pool, single_pool) kept — verifier-independent.
tests/inference_engine/scheduler/test_scheduler.py	−421	All 20 tests against DeterministicEngine. Migrated selectively to integration.

What was added

File	Lines	What
tests/integration/test_scheduler_real.py	+422, 12 tests	Scheduler runtime against real SpeculativeEngine over Qwen3-0.6B: construction validation, happy path, REJECT/QUEUE admission, cancellation, engine error propagation (parametric injector wrapping the real engine, same pattern PR-N1 used for gRPC error-mapping), 3-way concurrency, shutdown, active-count drain. Looser assertions than the originals (real engine output varies); structural invariants are what matters.
tests/integration/conftest.py	+82	New session-scoped real_speculative_engine fixture (Qwen3-0.6B + SparseLogitsProposer + SpeculativeDecoder + SpeculativeEngine wrapper). Mirrors tests/system/conftest.py but uses 0.6B not 1.7B to match the rest of the integration suite.
tests/inference_engine/scheduler/test_scheduler_validation.py	+102, 5 tests	Pre-engine validation paths on Linux: pool-dim mismatch, empty prompt, zero max_new_tokens, empty EOS. All run with engine=None; validation rejects before the scheduler enqueues the worker.
scripts/review_pr_n2_on_mac.sh	+88	Mac M4 reviewer aid.

CI workflow change

.github/workflows/ci.yaml: split --cov=inference_engine.scheduler into per-module coverage:

- --cov=inference_engine.scheduler
+ --cov includes:
+   inference_engine/scheduler/config.py           (Linux ✓)
+   inference_engine/scheduler/session.py          (Linux ✓)
+   inference_engine/scheduler/pooled_verifier.py  (Linux ✓ via exempt test)
+ --cov excludes (integration-only):
+   inference_engine/scheduler/scheduler.py

Also pre-emptively switched to coverage run -m pytest + coverage report --include (the same pattern Mac reviewer scripts use, to dodge the torch+pytest-cov race that hit PR-N1's CI runs).

Linux verification

PYTHONPATH=.:sdks/python coverage run -m pytest <Linux gate paths>:
  680 passed (was 695 on main; -15 net = removed 20 scheduler-engine tests,
              added 5 engine-independent validation tests).
  100% coverage on 1456 stmts (was 1660 on main; -204 = scheduler.scheduler now
              integration-only).

Mac M4 evidence (REQUIRED for merge)

Per ADR 0008 §9: scheduler runtime correctness now lives in the integration suite. Reviewer runs:

bash scripts/review_pr_n2_on_mac.sh
git add results/platform-tests/pr-n2-mac-*
git commit -m "Mac M4 review evidence for PR-N2"
git push

Acceptance: all integration tests pass against real Qwen3-0.6B, including PR-N1's coordinator/generator suites (if PR-N1 is also on the merge target) and the INV-3 byte-exact GA gate (PR-E1).

Stack

PR-N2 is branched off main, independent of PR-N1 (#53) at the file level. The two touch disjoint test files (session/ vs scheduler/) and disjoint conftest copies. Can merge in either order.

Once both land, PR-N3 cleans up the HTTP shim's doubles + subtypes (largest remaining cluster: tests/inference_engine/server/conftest.py + test_app_*.py + test_engine.py + test_tokenizer.py).