PR-N3: remove HTTP-shim, engine, tokenizer test doubles

[FluffyAIcode]

Why

Third installment of the no-test-doubles cleanup. Largest single PR by code volume — retires the entire cluster of HTTP-shim, SpeculativeEngine, tokenizer, and streaming-detokenizer test mirrors that lived in tests/inference_engine/server/.

Scope

Per the user's principle (fake = mock, all banned), PR-N3 deletes:

Linux test file	Δ	Test doubles removed
tests/inference_engine/server/conftest.py	−213 / +63	DeterministicEngine, DeterministicTokenizer + 3 fixtures
tests/inference_engine/server/test_app_routes.py	−334, 13 tests	(used the doubles)
tests/inference_engine/server/test_app_streaming.py	−524, 14 tests	(used the doubles)
tests/inference_engine/server/test_app_with_scheduler.py	−347, 12 tests	(used the doubles)
tests/inference_engine/server/test_app_metrics_and_auth.py	−496, 21 tests	_AlwaysHoldingEngine, _KVAwareSlowEngine, _ProxyEngine, _EmptyTemplateTokenizer
tests/inference_engine/server/test_engine.py	−325, 18 tests	_DecoderDouble, _DecoderResult, _VerifierDouble, _LegacyVerifierDouble
tests/inference_engine/server/test_tokenizer.py	−86, 6 tests	_BrokenTokenizer, _EmptyTemplateTokenizer, _NoEosTokenizer
tests/inference_engine/server/test_streaming.py	−62, 4 tests	(used DeterministicTokenizer._intern)

Total deletions: ~2,400 lines, 88 tests.

What was added

Integration file	Lines	Tests	Coverage focus
tests/integration/test_http_shim_real.py	228	10	Chat-completions OpenAI envelope, validation rejects, SSE streaming, auth (3 paths), public endpoints (healthz/metrics), /v1/models
tests/integration/test_engine_real.py	124	7	SpeculativeEngine wrapper: tokenizer/model_id_label, EngineResult shape, max_new_tokens honored, error rejections, on_token callback (per-token + early-stop)
tests/integration/test_tokenizer_real.py	98	6	Real Qwen3 tokenizer satisfies Tokenizer Protocol; resolve_eos_ids includes canonical EOS + `<
tests/integration/test_streaming_real.py	66	4	_StreamingDetokenizer per-token deltas + per-instance state + special-token handling

Linux file (no doubles)	Lines	Tests	Coverage focus
tests/inference_engine/server/test_errors.py	+40	2	request_validation_exception_handler single-error path; unhandled_exception_handler 500 envelope (paths previously only exercised via the deleted HTTP-shim tests)

Reviewer aid	Lines
scripts/review_pr_n3_on_mac.sh	93

CI workflow change

.github/workflows/ci.yaml — Linux gate --include filter dropped:

• inference_engine.server.app
• inference_engine.server.engine
• inference_engine.server.tokenizer
• inference_engine.server.streaming

Linux gate now covers ONLY:

• inference_engine.server.{auth, config, errors, grpc_app, metrics, schemas, proto_gen}
• inference_engine.memory.*
• inference_engine.scheduler.{config, session, pooled_verifier}
• inference_engine.pipeline.*
• inference_engine.session.store
• sdks.python.kakeya.*
• training.repr_align.*

Linux verification

PYTHONPATH=.:sdks/python coverage run -m pytest <Linux gate paths>:
  609 passed, 100% coverage on 1148 stmts.
  (was 695 / 1660 on main; -86 tests = removed 88 HTTP-shim tests, added 2 errors-handler tests.
   -512 stmts = server.app + server.engine + server.tokenizer + server.streaming now integration-only.)

Mac M4 evidence (REQUIRED for merge)

Per ADR 0008 §9 — bash scripts/review_pr_n3_on_mac.sh runs the full integration suite (PR-E1 INV-3 + PR-N1 coordinator/generator + PR-N2 scheduler + PR-N3 http_shim/engine/tokenizer/streaming) against real Qwen3-0.6B and produces pr-n3-mac-integration-tests-<unix>.json evidence.

Stack

PR-N3 is branched off main, independent of PR-N1 (#53) and PR-N2 (#54) at the file level. The three touch disjoint test files. Can merge in any order.

Next PR

PR-N4 (final): SDK conftest stub (_MinimalVerifierStub) cleanup + final CI workflow consolidation. Smaller scope; closes the no-test-doubles cleanup sequence.