test(orchestration): multi-round Phase 3 output-channel regression

[KodBena]

Summary

Adds a regression test that pins the post-v1.0.27 contract at the
adaptive_reevaluate multi-round Phase 3 path specifically.

Companion to the framework-level regression at
tests/test_orchestration_middleware.py::TestTrailingYieldAfterSpawnPrimitive::test_trailing_yield_after_parallel_reaches
(which pins the abstract trailing-yield contract via ctx.parallel).
This new test pins the adaptive_reevaluate-specific manifestation:
multi-round dispatch with max_rounds>=2 must deliver every round's
Stage 2 previews AND the Stage 3 finalization stage to the wire via
caps.send_response.

Discrimination confirmed end-to-end

Test was verified to FAIL on the actual pre-v1.0.27 substrate by
restoring middleware/orchestration.py from commit d918bf8 in a
git worktree (keeping the v1.0.27 test file + harness in place), then
running the test. Failure message:

AssertionError: Stage 3 stranding regression: expected one
authoritative per analyzed turn ([0, 1, 2, 3, 4, 5]); got []. 
Pre-v1.0.27 the drain race stranded these in ctx._output_queue 
with no future handle_response to drain them.

Passes on main (the v1.0.27 substrate).

Three assertions

1. Exactly one is_during_search=False per analyzed turn — Stage 3
   finalization reaches the wire.
2. Total preview count ≥ N_TURNS + sum(per-round spawn counts) —
   every round's Stage 2 previews reach, not just round 1's.
3. Total emission count ≥ multi-round lower bound — not the
   single-round-only stranding shape.

Test plan

• Test passes on current main substrate (486 pass, 0 fail).
• Test fails on actual pre-v1.0.27 substrate via git worktree
  (verified manually; failure message names the stranding
  mechanism explicitly).
• mypy --strict clean on the touched test file.
• No production code change — test-only addition.

🤖 Generated with Claude Code