PR-E1 (ADR 0008 §6.5): integration suite + INV-3 byte-exact GA gate

[FluffyAIcode]

What ships

Per ADR 0008 §6.5, PR-E1 delivers:

1. tests/integration/ directory

• __init__.py
• conftest.py — auto-applies @pytest.mark.integration to every test in the directory; bare pytest skips them, contributors opt in via pytest -m integration.

2. tests/integration/test_inv3_session_determinism_gate.py

The INV-3 byte-exact GA gate (ADR 0008 §7 G3). Drives two independent SinkWindowVerifier instances (real Qwen3-0.6B weights via fresh_verifier_factory) through identical history fed via different chunkings, asserts the resulting greedy token streams are byte-identical. Three tests:

Test	What it covers
test_one_call_vs_two_calls_yield_byte_identical_tokens	Minimal: 1×10 vs 2×5 chunking, 12 tokens generated greedy.
test_chunking_invariance_across_three_splits	Stronger: 1×20 / 3×medium / 10×small chunkings, 8 tokens generated. Catches chunk-boundary bugs the 1-vs-2 case might miss (e.g., a chunk crossing a sink+window trim boundary).
test_repeated_runs_with_same_history_byte_identical	Sanity: same workload twice on same verifier == identical output.

This replaces tests/core/test_determinism_gate.py (deleted in PR-A3 along with verifier.path_select). Per ADR 0008 §6.6, the replacement lives in tests/integration/ rather than tests/core/ because integration is where Mac-M4-only GA gates belong per §9.

3. pytest.ini

Minimal new file registering the integration marker so opt-in invocations (pytest -m integration) don't trigger PytestUnknownMarkWarning.

4. scripts/review_pr_e1_on_mac.sh

Mac M4 reviewer aid. Runs pytest -m integration tests/integration/ and produces pr-e1-mac-integration-tests-<unix>.json under results/platform-tests/. Same coverage-free pattern as review_pr_b3_on_mac.sh.

Independence from PR-D1

PR-E1 was originally stacked on PR-D1 (#49) but reviewed-once it's clear the two are file-disjoint: PR-E1 only adds new files under tests/integration/, plus pytest.ini and scripts/review_pr_e1_on_mac.sh. It does not depend on PR-D1's deletions. Rebased onto main directly so CI triggers normally.

The two PRs can merge in either order.

Not in this PR (deferred)

• scripts/bench_agentic/bench_session_long_run.py: §6.5 also mentions a Mac M4 long-session bench using the gRPC SDK. Splitting it out so PR-E1's diff stays focused on the GA gate. Will land as PR-E1b or rolled into PR-E2.
• PR-E2: GitHub Actions self-hosted Mac M4 runner workflow invoking pytest -m integration on every PR labelled needs-mac-m4. Until that workflow lands, the gate runs manually via scripts/review_pr_e1_on_mac.sh.

Linux verification

Linux CI gate (existing 8 test paths):
  682 passed, 100% coverage  ← UNCHANGED. tests/integration/ is not in the Linux paths.

Integration suite collection:
  3 tests collected from tests/integration/, marker auto-applied via conftest.

Mac M4 evidence (REQUIRED for merge — load-bearing for v0.3 GA)

Per ADR 0008 §9, this PR's true validation happens on Mac M4. Linux CI cannot validate INV-3 against real Qwen3 numerics. Reviewer runs:

bash scripts/review_pr_e1_on_mac.sh
git add results/platform-tests/pr-e1-mac-*
git commit -m "Mac M4 review evidence for PR-E1"
git push

…and pushes the JSON evidence to this PR branch before merge. All 3 tests must pass with byte-exact equality. Any failure here means INV-3 is broken on real numerics → BLOCKS v0.3 GA.

Next PR after merge

• PR-E1b or fold-in: bench_session_long_run.py against the gRPC SDK.
• PR-E2: self-hosted Mac M4 GitHub Actions workflow.
• PR-D2 (independent): HTTP shim refactor onto SessionStore.